<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[CPC AI — Solo MVP Build Log]]></title><description><![CDATA[CPC AI — Solo MVP Build Log]]></description><link>https://cpcai.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 19:00:16 GMT</lastBuildDate><atom:link href="https://cpcai.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[アプリ開発の見積もりが10倍違う理由。3〜5日でMVPを出すスタジオの話]]></title><description><![CDATA[「予算は150万円です。3社に見積もってもらったら、全部400万円以上で、納期は4か月でした。なぜこんなに違うんでしょうか」 この質問は、値段の話に見えて、実は「あなたとコードを書く人の間に何人いるか」という話です。 私はソウルにある1人だけの開発スタジオ、CPC AI Inc…]]></description><link>https://cpcai.hashnode.dev/10-3-5-mvp</link><guid isPermaLink="true">https://cpcai.hashnode.dev/10-3-5-mvp</guid><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:35:38 GMT</pubDate><content:encoded><![CDATA[<p>「予算は150万円です。3社に見積もってもらったら、全部400万円以上で、納期は4か月でした。なぜこんなに違うんでしょうか」</p>
<p>この質問は、値段の話に見えて、実は「あなたとコードを書く人の間に何人いるか」という話です。</p>
<p>私はソウルにある1人だけの開発スタジオ、CPC AI Inc. の唯一の開発者です。AIを使ってWebアプリとモバイルアプリを3〜5日で作っています。営業も、プロジェクトマネージャーも、デザイン会社も、下請けもいません。だから価格の桁が変わり、だから納期が「か月」ではなく「日」で数えられます。</p>
<h2>まず、みんなが最初に聞く数字</h2>
<p>予約と決済が入ったアプリ（カレンダー、空き枠、カード決済、通知、管理画面）を同じ範囲で作る場合の相場です。</p>
<ul>
<li><strong>一般的な制作会社：</strong> 250万〜600万円、3〜6か月、4〜8人のチーム</li>
<li><strong>フリーランスのシニア：</strong> 60万〜200万円、1〜2.5か月、1人だが掛け持ち</li>
<li><strong>直接作る1人スタジオ：</strong> 15万円前後から、機能の芯の部分は3〜5日
私のサイトでは、同じ範囲の一般的な見積もりと比較して「最大20分の1の価格」と書いています。この数字を正直に読み替えると、「5日で作るアプリは原価が安い」ではなく、「制作会社の見積もりには、作っていない人がたくさん含まれている」ということです。その層を外すと、残るのは実際の開発時間だけです。予約アプリの実際の開発時間は、月ではなく日です。</li>
</ul>
<h2>見積もりを本当に動かす3つのこと</h2>
<p><strong>1. 意思決定が何回、あなたの側にあるか。</strong> 質問に2時間で返ってくる案件は、2日間「料金はレッスン単位ですか、人単位ですか？」が宙に浮く案件より5倍速く進みます。この沈黙は見積もりには出ませんが、カレンダーには出ます。</p>
<p><strong>2. アカウントが誰の名義か。</strong> ドメイン、データベース、決済代行、ストアの登録。制作会社の名義になっている場合、それを直すのが最初の作業になります。</p>
<p><strong>3. 「ほぼ完成」から「公開」までの距離。</strong> Lovable や Cursor で80%まで作ったアプリは、残り20%の作業ではありません。別の仕事です。権限の見直し、Webhookが二重に届いても二重課金しない仕組み、本番環境とテスト環境の分離。</p>
<h2>3〜5日のMVPとは、具体的に何が出てくるのか</h2>
<p>デモではありません。スマートフォンで開いて、実際のデータで使えるアプリです。クライアント案件での1日ずつの流れです。</p>
<p><strong>1日目 — 1本の流れが、実データで通る。</strong> アプリ全体ではありません。会員登録、メイン画面、再読み込みしても残る書き込み。見た目はあえて地味にします。1日目にデザインを詰めると1日かかり、5日目なら1時間で済むからです。</p>
<p><strong>2日目 — お金が、1回だけ、正しく動く。</strong> 注文の作成、サーバー側での決済確認、Webhook、返金の経路。その日の終わりのテストはシンプルです。実際の支払い1回で、権利が1つだけ付与される。Webhookが2回届いても、二重課金にも二重付与にもならない。</p>
<p><strong>3日目 — 2人目のユーザーに耐える継ぎ目。</strong> アカウントAがURLのIDを書き換えてBのデータを読もうとする。返ってくるなら、バックエンド全体に権限チェックを入れ直します。
<strong>4日目 — あなたのドメインで、あなた以外の人が使える状態に。</strong> 秘密情報をリポジトリから出す、ゼロからデプロイ、証明書、エラー監視、一度は復元してみたバックアップ。</p>
<p><strong>5日目 — あなたが使って、私が見つかった不具合を直す。</strong> ここで仕様書に書かれていなかった事実が出てきます。現場はいつもタブレットから予約している、講座名にスラッシュが入っていてURLが壊れる、など。</p>
<p>そして<strong>含まれないもの</strong>は、着手前に書面で出します。レビュー機能、クーポン、多言語、予約配信、Excelからのデータ移行。除外リストのない見積もりは、範囲ではなく願望です。</p>
<h2>契約と請求書 — 地味だが、すべてを決める部分</h2>
<p>安い制作会社が安くなくなる瞬間は、経理から「請求書をください」と言われて、出せるものがないときです。技術的には良い案件が、そこだけで数か月止まるのを何度も見ました。海外の匿名アカウントに発注していた、というだけの理由で。
私の条件は、支払い前に確認できます。</p>
<ul>
<li><strong>契約前は0円。</strong> プロジェクトの確認、見積もり、計画は無料です。「これは3週間必要です。理由はこうです」と答える場合、あなたが失うのは1時間で、手付金ではありません。</li>
<li><strong>動くものを見てから残金。</strong> 契約時に手付金、残りは上記の日数が過ぎて、あなたが触ってから。</li>
<li><strong>韓国の法人との正式契約、正式な請求書。</strong> CPC AI Inc.、法人番号 771-88-03539、ソウル。フリーランス仲介サイトの匿名プロフィールではありません。</li>
<li><strong>アカウントは最初から、あなたの名義で。</strong> データベースを制作会社の名義で持つスタジオは、やさしく言って製品を人質に取っています。</li>
</ul>
<h2>よくある質問</h2>
<p><strong>予約アプリの開発費はいくらですか。</strong> カレンダー、空き枠、決済、通知、管理画面という芯の部分で、15万円前後からです。見積もりは無料で、範囲を書面でお出しします。</p>
<p><strong>韓国の開発者に頼むのは不安です。</strong> 不安の元は国ではなく契約です。法人番号、正式な請求書、アカウントの名義、そして書面の完成条件。この4つがあれば、距離は関係ありません。</p>
<p><strong>やり取りはどの言語ですか。</strong> 文章で、あなたの言語で行います。メールとチャットで、通話は必須ではありません。ソウルと日本は時差がなく、日本語でそのまま書いていただいて構いません。</p>
<p><strong>仕上がりが気に入らなかったら。</strong> 納得いくまで修正は含まれています。残金は、動くアプリを見てからです。</p>
<p><strong>App Store / Google Play で公開できますか。</strong> できます。ただしストアの審査は開発日数ではありません。5日目に公開できる状態になりますが、公開はストアの時計で進みます。</p>
<hr />
<p>アイデアがある、プロトタイプが80%で止まっている、見積もりが合わない。どれでも、支払いの前に「どの種類の仕事か」をお伝えします。</p>
<p><strong><a href="https://cpcai-studio.netlify.app/ja/">https://cpcai-studio.netlify.app/ja/</a></strong> — 確認は無料、その後で範囲と価格を確定。
<strong><a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></strong></p>
<p>CPC AI Inc. · 韓国・ソウル · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[2026'da uygulama yaptırmanın gerçek maliyeti (ve aynı işin neden 20 kat fark ettiği)]]></title><description><![CDATA[Neredeyse her hafta aynı mesaj geliyor, kelimeler değişse de: "Bir serbest çalışandan 8.000 dolar teklif aldım, iki ajans 40.000 dolar ve üç ay dedi. Aynı…]]></description><link>https://cpcai.hashnode.dev/uygulama-gelistirme-maliyeti-2026</link><guid isPermaLink="true">https://cpcai.hashnode.dev/uygulama-gelistirme-maliyeti-2026</guid><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:32:39 GMT</pubDate><content:encoded><![CDATA[<p>Neredeyse her hafta aynı mesaj geliyor, kelimeler değişse de: "Bir serbest çalışandan 8.000 dolar teklif aldım, iki ajans 40.000 dolar ve üç ay dedi. Aynı uygulama, aynı ekranlar. Bu nasıl oluyor?"</p>
<p>Bu bir fiyat sorusu gibi görünüyor. Aslında şu soru: sizinle kodu yazan kişi arasında kaç insan var?</p>
<p>Tek başıma çalışıyorum. Seul'deki tek kişilik bir stüdyo olan CPC AI Inc.'in tek geliştiricisiyim ve yapay zekâ ile web ve mobil uygulamaları 3-5 günde çıkarıyorum. Satış temsilcisi, proje yöneticisi, tasarım ajansı, taşeron yok. Fiyat farkını bu açıklıyor; sürenin de neden "çeyrek" değil "gün" ile ölçüldüğünü bu açıklıyor.</p>
<h2>Herkesin ilk sorduğu sayı</h2>
<p>Aynı rezervasyon uygulaması için (takvim, kontenjan, kartla ödeme, bildirimler, yönetim paneli):</p>
<ul>
<li><strong>Klasik ajans:</strong> 15.000-40.000 dolar, 3-6 ay, 4-8 kişilik ekip.</li>
<li><strong>Kıdemli serbest çalışan:</strong> 4.000-12.000 dolar, 4-10 hafta, tek kişi ama takvimi paylaşımlı.</li>
<li><strong>Doğrudan üreten tek kişilik stüdyo:</strong> yaklaşık 1.000-1.500 dolardan başlar, çekirdek işlev 3-5 gün. Kendi sitem, aynı kapsam için tipik ajans teklifleriyle karşılaştırıp "fiyatın 1/20'sine kadar" diyor. Bu sayının dürüst okuması "beş günde uygulama üretmek ucuz" değil. Okuması şu: "bir ajans teklifinde, üretmeyen çok insan var." O katmanları çıkarınca geriye gerçek geliştirme süresi kalıyor; bir rezervasyon uygulamasının gerçek geliştirme süresi ise ay değil, gün.</li>
</ul>
<h2>Teklifi gerçekten değiştiren üç şey</h2>
<p><strong>1. Kaç karar sizde kalıyor.</strong> Müşterinin iki saatte cevap verdiği proje, "fiyat ders başına mı kişi başına mı?" sorusunun iki gün havada kaldığı projeden beş kat hızlı ilerliyor. Bu sessizlik teklifte görünmez, takvimde görünür.</p>
<p><strong>2. Hesaplar kimin adına.</strong> Alan adı, veritabanı, ödeme sağlayıcısı, App Store ve Google Play geliştirici hesabı. Stüdyo adına açılmışsa, bunları düzeltmek işin ilk maddesi olur, sonundaki ayrıntı değil.</p>
<p><strong>3. "Neredeyse bitti" ile "yayında" arasındaki mesafe.</strong> Lovable veya Cursor ile %80'e gelmiş bir uygulama, işin %20'si değildir. Başka bir iştir: yetkileri gözden geçirmek, tekrar eden webhook'u idempotent yapmak, üretim ortamını test ortamından ayırmak.</p>
<h2>3-5 günlük MVP'de tam olarak ne çıkıyor</h2>
<p>Demo değil: telefonunuzda açıp gerçek veriyle kullandığınız bir uygulama. Müşteri projelerinde gün gün:</p>
<p><strong>1. gün — tek bir akış, baştan sona, gerçek veriyle.</strong> Uygulamanın tamamı değil: kayıt, ana ekran, sayfa yenilenince kaybolmayan bir yazma işlemi. Bilerek sade görünüyor; birinci gün özenli tasarım bir gün, beşinci gün bir saat eder.</p>
<p><strong>2. gün — para bir kez, doğru şekilde hareket ediyor.</strong> Sipariş oluşturma, sunucu tarafında ödeme onayı, webhook, iade yolu. Akşamın testi: gerçek bir ödeme tam olarak bir erişim açıyor, bir kez.</p>
<p><strong>3. gün — ikinci kullanıcıyı taşıyan dikişler.</strong> A ve B olmak üzere iki hesap var: A hesabı, URL'deki bir numarayı değiştirerek B'nin kaydını okumaya çalışıyor. Veri dönüyorsa tüm arka uç yetki kontrolünden geçmeli.</p>
<p><strong>4. gün — sizin alan adınızda, siz olmayan biri için çalışıyor.</strong> Ortam değişkenleri depodan çıkmış, sıfırdan dağıtım, sertifika, hata izleme, bir kez geri yüklenmiş yedek.</p>
<p><strong>5. gün — siz kullanıyorsunuz, ben bulduğunuzu düzeltiyorum.</strong> Burada hiçbir dokümanda yazmayan iki gerçek çıkıyor: ekip her zaman tabletten rezervasyon yapıyor ya da ders adlarında URL'yi bozan bir eğik çizgi var. Ve <strong>dahil olmayanlar</strong>, başlamadan önce yazılı olarak: değerlendirmeler, kuponlar, çok dillilik, planlı kampanyalar, Excel'den veri taşıma. Hariç tutma listesi olmayan bir teklif kapsam değil, temennidir.</p>
<h2>Sözleşme ve fatura: her şeyi belirleyen sıkıcı kısım</h2>
<p>Ucuz bir stüdyo, muhasebe bir belge istediğinde ve ortada belge yokken ucuz olmaktan çıkar. Koşullarım ödeme yapmadan önce kontrol edilebilir:</p>
<ul>
<li><strong>Sözleşmeden önce 0 dolar.</strong> Projeyi incelemek, tahmin ve plan ücretsiz. Cevabım "bu üç hafta ister, nedeni şu" ise bir saat kaybettiniz, bir ön ödeme değil.</li>
<li><strong>Kalan ödeme, çalışan yazılımı gördükten sonra.</strong> Sözleşme imzasında ön ödeme, gerisi yukarıdaki günler geçtikten ve siz tıkladıktan sonra.</li>
<li><strong>Gerçek bir Kore şirketiyle sözleşme ve resmi fatura.</strong> CPC AI Inc., ticaret sicil numarası 771-88-03539, Seul.</li>
<li><strong>Hesaplar baştan itibaren sizin adınıza.</strong></li>
</ul>
<h2>Sık sorulanlar</h2>
<p><strong>Bir rezervasyon uygulaması ne kadar?</strong> çekirdek işlev (takvim, kontenjan, ödeme, bildirim, panel) bende yaklaşık 1.000-1.500 dolardan başlıyor. Teklif ücretsiz ve yazılı kapsamla geliyor.</p>
<p><strong>Kore'deki bir geliştiriciyle çalışmak riskli mi?</strong> Risk ülkede değil, sözleşmede. Ticaret sicil numarası, resmi fatura, hesap sahipliği ve yazılı bitiş çizgisi isteyin.</p>
<p><strong>Hangi dilde çalışıyoruz?</strong> Yazılı olarak, sizin dilinizde. E-posta ve sohbet; zorunlu görüşme yok. Seul ile İstanbul arasında altı saat fark var ve görüşme yalnızca işi uzatır.</p>
<p><strong>Sonuç hoşuma gitmezse?</strong> Revizyonlar istediğiniz hâle gelene kadar dahil, kalan ödeme uygulama çalıştıktan sonra.</p>
<p>Bir fikriniz, tıkanmış bir prototipiniz ya da yetişmeyen bir teklifiniz mi var? Yazın; ödeme yapmadan önce ne tür bir işin içinde olduğunuzu söyleyeyim:</p>
<p><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a> — ücretsiz inceleme, sonra sabit kapsam ve fiyat. <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></p>
<p>CPC AI Inc. · Seul, Güney Kore · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[Quanto costa davvero sviluppare un'app nel 2026 (e perché lo stesso progetto costa 20 volte tanto)]]></title><description><![CDATA[La richiesta che arriva quasi ogni settimana suona sempre uguale: «Ho un preventivo di 8.000 euro da un freelance e di 45.000 da due agenzie. Stessa app…]]></description><link>https://cpcai.hashnode.dev/quanto-costa-davvero-sviluppare-un-app-nel-2026-e-perch-lo-stesso-progetto-costa-20-volte-tanto</link><guid isPermaLink="true">https://cpcai.hashnode.dev/quanto-costa-davvero-sviluppare-un-app-nel-2026-e-perch-lo-stesso-progetto-costa-20-volte-tanto</guid><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:30:45 GMT</pubDate><content:encoded><![CDATA[<p>La richiesta che arriva quasi ogni settimana suona sempre uguale: «Ho un preventivo di 8.000 euro da un freelance e di 45.000 da due agenzie. Stessa app, stesse schermate. Come è possibile?»</p>
<p>Non è una domanda sul prezzo. è una domanda su quante persone stanno tra te e chi scrive il codice.</p>
<p>Lavoro da solo. Sono l'unico sviluppatore di CPC AI Inc., uno studio di una persona a Seul, e costruisco app web e mobile con l'AI in 3-5 giorni. Nessun commerciale, nessun project manager, nessuna agenzia di design, nessun subappaltatore. Questo spiega la differenza di prezzo e spiega anche perché i tempi si contano in giorni e non in trimestri.</p>
<h2>Il numero che tutti chiedono per primo</h2>
<p>Per la stessa app di prenotazioni con pagamenti (calendario, posti disponibili, carta di credito, notifiche, pannello di amministrazione):</p>
<ul>
<li><strong>Agenzia tradizionale:</strong> 15.000-45.000 euro, 3-6 mesi, squadra di 4-8 persone.</li>
<li><strong>Freelance senior:</strong> 4.000-12.000 euro, 4-10 settimane, una persona ma con l'agenda condivisa.</li>
<li><strong>Studio di una persona che costruisce in prima persona:</strong> da circa 1.000-1.500 euro, 3-5 giorni per il nucleo funzionante.
Il mio sito confronta i preventivi tipici delle agenzie per lo stesso perimetro e arriva a «fino a 1/20 del prezzo». La lettura onesta non è «un'app in cinque giorni costa poco produrla». è: «in un preventivo d'agenzia c'è molta gente che non costruisce». Togliendo quei livelli resta il tempo reale di sviluppo.</li>
</ul>
<h2>Cosa fa davvero variare il preventivo</h2>
<p><strong>1. Quante decisioni restano a te.</strong> Un progetto in cui il cliente risponde in due ore va cinque volte più veloce di uno in cui la domanda «il prezzo è per lezione o per persona?» resta ferma due giorni. Quel silenzio non appare nel preventivo, ma appare nel calendario.</p>
<p><strong>2. A chi intestano gli account.</strong> Dominio, database, provider di pagamento, account sviluppatore Apple e Google. Se sono intestati allo studio, sistemarli diventa il primo punto di lavoro.</p>
<p><strong>3. Quanto dista «quasi finito» da «online».</strong> Un'app costruita con Lovable o Cursor all'80% non è il 20% del lavoro. è un lavoro diverso: controllare i permessi, rendere idempotente il webhook, separare produzione e ambiente di test.</p>
<h2>Cosa esce davvero da un MVP in 3-5 giorni</h2>
<p>Non una demo: un'app che apri sul telefono e usi con dati veri. Giorno per giorno, nei progetti dei clienti:</p>
<p><strong>Giorno 1 — un flusso completo, con dati reali.</strong> Non tutta l'app: registrazione, schermata principale, una scrittura che sopravvive al ricaricamento. è volutamente sobrio; il design curato costa un giorno il primo giorno e un'ora il quinto.</p>
<p><strong>Giorno 2 — i soldi si muovono, una volta, correttamente.</strong> Creazione dell'ordine, conferma lato server, webhook, percorso di rimborso. Il test della sera: un pagamento reale dà esattamente un accesso, una volta sola.
<strong>Giorno 3 — le giunture che reggono un secondo utente.</strong> Due account: l'account A prova a leggere il record di B cambiando un numero nell'URL. Se i dati tornano, tutto il backend va ripassato sui permessi.</p>
<p><strong>Giorno 4 — gira sul tuo dominio, per qualcuno che non sei tu.</strong> Variabili d'ambiente fuori dal repository, deploy da zero, certificato, monitoraggio errori, un backup che qualcuno ha già ripristinato.</p>
<p><strong>Giorno 5 — lo usi tu, io correggo quello che trovi.</strong> Qui emergono i due fatti che nessun documento aveva previsto: il team prenota sempre dal tablet, oppure il nome del corso contiene una barra che rompe l'URL.</p>
<p>E quello che <strong>non</strong> è incluso, per iscritto prima di iniziare: recensioni, coupon, multilingua, campagne programmate, migrazione da Excel. Un preventivo senza esclusione esplicita non è un perimetro: è una speranza.</p>
<h2>Contratto e fattura: la parte noiosa che decide tutto</h2>
<p>Uno studio economico smette di esserlo il giorno in cui l'amministrazione chiede un documento e non esiste. Le mie condizioni si verificano prima di pagare qualsiasi cosa:</p>
<ul>
<li><strong>0 euro prima del contratto.</strong> Analisi, stima e piano sono gratuiti. Se la risposta è «servono tre settimane, ed ecco perché», hai perso un'ora, non un acconto.</li>
<li><strong>Saldo dopo aver visto il software funzionare.</strong> Acconto alla firma, il resto quando i giorni sopra sono già passati e li hai cliccati.</li>
<li><strong>Contratto con una società coreana reale e fattura ufficiale.</strong> CPC AI Inc., numero di registrazione 771-88-03539, Seul.</li>
<li><strong>I tuoi account, a tuo nome, dal primo giorno.</strong></li>
</ul>
<h2>Domande frequenti</h2>
<p><strong>Quanto costa un'app di prenotazioni?</strong> Il nucleo funzionante (calendario, posti, pagamento, notifiche, pannello) parte da circa 1.000-1.500 euro. Il preventivo è gratuito e arriva con il perimetro scritto.</p>
<p><strong>è rischioso lavorare con uno sviluppatore in Corea?</strong> Il rischio non è il paese, è il contratto. Chiedi numero di registrazione, fattura ufficiale, proprietà degli account e una linea di arrivo scritta.</p>
<p><strong>In che lingua lavoriamo?</strong> Per iscritto, nella tua lingua. Email e chat, senza call obbligatorie: Seul e l'Italia hanno otto ore di differenza e una call allunga solo le cose.</p>
<p><strong>Pubblicate su App Store e Google Play?</strong> Sì, ma la revisione dello store non è un giorno di sviluppo: l'app può essere pronta il giorno 5, la pubblicazione segue l'orologio dello store.</p>
<p>Hai un'idea, un prototipo bloccato o un preventivo che non torna? Scrivi e ti dico in che tipo di lavoro ti trovi prima di qualsiasi pagamento:</p>
<p><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a> — analisi gratuita, poi perimetro e prezzo fissi.
<a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></p>
<p>CPC AI Inc. · Seul, Corea del Sud · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[Freelance ou studio : les 7 questions à poser avant de payer (checklist)]]></title><description><![CDATA[Avant de verser un acompte pour une application, il y a une demi-heure de questions qui évite trois mois de problèmes. Je les pose moi-même à l'envers…]]></description><link>https://cpcai.hashnode.dev/freelance-ou-studio-les-7-questions-poser-avant-de-payer-checklist</link><guid isPermaLink="true">https://cpcai.hashnode.dev/freelance-ou-studio-les-7-questions-poser-avant-de-payer-checklist</guid><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:27:56 GMT</pubDate><content:encoded><![CDATA[<p>Avant de verser un acompte pour une application, il y a une demi-heure de questions qui évite trois mois de problèmes. Je les pose moi-même à l'envers, parce que je suis de l'autre côté : studio d'une personne à Séoul, applications livrées en 3 à 5 jours, jusqu'à 1/20 du prix d'un devis d'agence classique.</p>
<p>Voici la liste, dans l'ordre où elle sert.</p>
<p><strong>1. « Qui possède les comptes ? »</strong>
Domaine, base de données, prestataire de paiement, compte développeur Apple et Google. La bonne réponse est : vous, dès le premier jour. Un prestataire qui garde la base de données à son nom ne vous bloque pas par malveillance ; il vous bloque parce que partir devient coûteux.</p>
<p><strong>2. « Que se passe-t-il si je pars ? »</strong>
Demandez le plan de sortie avant de commencer : où sont les secrets, comment on restaure une sauvegarde, qui peut déployer. Si la réponse est vague, c'est déjà la réponse.</p>
<p><strong>3. « Qu'est-ce qui n'est pas inclus ? »</strong>
Un devis sans liste d'exclusions n'est pas un périmètre. Avis, coupons, multilingue, campagnes programmées, migration depuis un tableur : chaque ligne oubliée revient plus tard en avenant.</p>
<p><strong>4. « Combien de décisions m'attendront, et sous quel délai ? »</strong>
Le vrai facteur de délai n'est pas la technologie, c'est la boucle de décision. « Le tarif est par cours ou par personne ? » qui reste deux jours sans réponse coûte plus qu'un framework mal choisi.
<strong>5. « Comment le paiement est-il confirmé ? »</strong>
Réponse attendue : côté serveur, avec un webhook idempotent. Si la confirmation se fait sur l'écran de succès, l'application est offerte à qui tape la bonne URL.</p>
<p><strong>6. « Puis-je tester avec deux comptes ? »</strong>
C'est le test qui compte : le compte A essaie de lire l'enregistrement de B en changeant un identifiant. Presque tous les prototypes générés par IA échouent ici, parce qu'ils ont été testés avec un seul compte.</p>
<p><strong>7. « Quelle facture vais-je recevoir ? »</strong>
Une vraie société, un numéro d'immatriculation, une facture officielle. Chez moi : CPC AI Inc., Séoul, 771-88-03539, contrat avant tout paiement, solde après avoir vu l'application tourner.</p>
<p>Si les sept réponses tiennent debout, le reste n'est plus qu'une question de prix. Sinon, aucun tarif n'est assez bas.</p>
<p><strong>écrire ne coûte rien et ne vous engage à rien :</strong>
<a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a> — analyse gratuite, périmètre et prix fermes ensuite.
<a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></p>
<p>CPC AI Inc. · Séoul, Corée du Sud · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[App entwickeln lassen: warum der Preis um Faktor 20 schwankt – und was in 5 Tagen fertig ist]]></title><description><![CDATA[Der Satz kommt in fast jedem ersten Gespräch: „Ich habe 10.000 Euro. Zwei Agenturen wollen 45.000 und vier Monate. Wie kann das sein?“ Es ist keine…]]></description><link>https://cpcai.hashnode.dev/app-entwickeln-lassen-warum-der-preis-um-faktor-20-schwankt-und-was-in-5-tagen-fertig-ist</link><guid isPermaLink="true">https://cpcai.hashnode.dev/app-entwickeln-lassen-warum-der-preis-um-faktor-20-schwankt-und-was-in-5-tagen-fertig-ist</guid><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:19:00 GMT</pubDate><content:encoded><![CDATA[<p>Der Satz kommt in fast jedem ersten Gespräch: „Ich habe 10.000 Euro. Zwei Agenturen wollen 45.000 und vier Monate. Wie kann das sein?“</p>
<p>Es ist keine Preisfrage. Es ist eine Frage danach, wie viele Menschen zwischen Ihnen und der Person sitzen, die den Code schreibt.</p>
<p>Ich arbeite allein. Ich bin der einzige Entwickler von CPC AI Inc., einem Ein-Personen-Studio in Seoul, und ich baue Web- und Mobile-Apps mit KI in 3 bis 5 Tagen. Kein Vertrieb, kein Projektmanager, keine Designagentur, keine Subunternehmer. Das erklärt den Preisunterschied – und es erklärt, warum die Zeit in Tagen gemessen wird und nicht in Quartalen.</p>
<h2>Die Zahl, nach der alle zuerst fragen</h2>
<p>Für dieselbe Buchungs-App mit Zahlung (Kalender, freie Plätze, Kartenzahlung, Benachrichtigungen, Admin-Bereich):</p>
<ul>
<li><strong>Klassische Agentur:</strong> 15.000 bis 45.000 Euro, 3 bis 6 Monate, Team von 4 bis 8 Personen.</li>
<li><strong>Senior-Freelancer:</strong> 4.000 bis 12.000 Euro, 4 bis 10 Wochen, eine Person mit geteiltem Kalender.</li>
<li><strong>Ein-Personen-Studio, das direkt baut:</strong> ab etwa 1.000 bis 1.500 Euro, 3 bis 5 Tage für den funktionierenden Kern.
Meine eigene Seite vergleicht mit üblichen Agenturangeboten für denselben Umfang und nennt bis zu 1/20 des Preises. Die ehrliche Lesart dieser Zahl ist nicht „eine App in fünf Tagen ist billig zu produzieren“. Sie ist: „In einem Agenturangebot steckt viel Personal, das nicht baut.“ Nimmt man diese Schichten weg, bleibt echte Entwicklungszeit – und die echte Entwicklungszeit einer Buchungs-App sind Tage.</li>
</ul>
<h2>Was den Preis wirklich bewegt</h2>
<p><strong>1. Wie viele Entscheidungen bei Ihnen liegen.</strong> Ein Projekt, in dem der Kunde in zwei Stunden antwortet, läuft fünfmal schneller als eines, in dem die Frage „gilt der Preis pro Kurs oder pro Person?“ zwei Tage offen bleibt. Diese Stille steht in keinem Angebot, aber sie steht im Kalender.</p>
<p><strong>2. Auf wessen Namen die Konten laufen.</strong> Domain, Datenbank, Zahlungsdienstleister, Store-Eintrag. Laufen sie auf den Namen des Studios, ist das Aufräumen der erste Arbeitspunkt – nicht ein Detail am Ende.</p>
<p><strong>3. Wie weit „fast fertig“ von „live“ entfernt ist.</strong> Eine mit Lovable oder Cursor gebaute App bei 80 Prozent ist keine 20 Prozent Arbeit. Es ist eine andere Arbeit: Berechtigungen prüfen, doppelt eintreffende Webhooks entschärfen, Produktion von Testumgebung trennen.</p>
<h2>Was ein 3-bis-5-Tage-MVP konkret ist</h2>
<p>Keine Demo. Eine App, die Sie auf dem Handy öffnen und mit echten Daten benutzen. Tag für Tag, wie es in Kundenprojekten läuft:</p>
<p><strong>Tag 1 – ein Ablauf, vollständig, mit echten Daten.</strong> Nicht die ganze App: Registrierung, der Hauptbildschirm, ein Schreibvorgang, der ein Neuladen übersteht. Es sieht absichtlich schlicht aus; schönes Design kostet an Tag 1 einen Tag, an Tag 5 eine Stunde.</p>
<p><strong>Tag 2 – Geld bewegt sich, einmal, korrekt.</strong> Bestellung anlegen, Zahlung serverseitig bestätigen, Webhook, Erstattungsweg. Der Test am Abend: Eine echte Zahlung schaltet genau einen Zugang frei, einmal. Kommt der Webhook zweimal, zahlt der Kunde nicht doppelt.</p>
<p><strong>Tag 3 – die Nahtstellen, die einen zweiten Nutzer aushalten.</strong> Zwei Konten: Konto A versucht, den Datensatz von B zu lesen, indem es eine ID in der URL ändert. Kommen Daten zurück, braucht das gesamte Backend einen Berechtigungsdurchlauf.
<strong>Tag 4 – es läuft auf Ihrer Domain, für jemanden, der nicht Sie ist.</strong> Secrets aus dem Repository, Deployment von Null, Zertifikat, Fehler-Monitoring, ein Backup, das schon einmal zurückgespielt wurde.</p>
<p><strong>Tag 5 – Sie benutzen es, ich korrigiere, was Sie finden.</strong> Hier tauchen die zwei Fakten auf, die keine Spezifikation enthielt: Ihr Team bucht immer vom Tablet, oder in den Kursnamen steckt ein Schrägstrich, der die URL zerlegt.</p>
<p>Und was <strong>nicht</strong> dazugehört, schriftlich vor dem Start: Bewertungen, Gutscheine, Mehrsprachigkeit, geplante Kampagnen, Migration aus Excel. Ein Angebot ohne Ausschlussliste ist kein Umfang, sondern eine Hoffnung.</p>
<h2>Vertrag und Rechnung – der langweilige Teil, der alles entscheidet</h2>
<p>Ein billiges Studio ist nicht mehr billig, sobald die Buchhaltung ein Dokument verlangt und keines existiert. Ich habe genau diesen Fall bei Kunden erlebt: technisch gutes Projekt, monatelang blockiert, weil der Anbieter kein Unternehmen hinter sich hatte.
Meine Bedingungen sind prüfbar, bevor Sie etwas zahlen:</p>
<ul>
<li><strong>0 Euro vor dem Vertrag.</strong> Projektsichtung, Einschätzung und Angebot sind kostenlos. Wenn meine Antwort „das braucht drei Wochen, und hier ist warum“ lautet, haben Sie eine Stunde verloren, keine Anzahlung.</li>
<li><strong>Rest nach sichtbarer Software.</strong> Anzahlung bei Vertragsschluss, der Rest, nachdem die Tage oben passiert und von Ihnen angeklickt sind.</li>
<li><strong>Vertrag mit einer echten koreanischen Gesellschaft und offizielle Rechnung.</strong> CPC AI Inc., Handelsregisternummer 771-88-03539, Seoul. Ausstellbar für Ihre Buchhaltung, mit den Umsatzsteuer-Angaben, die Sie brauchen.</li>
<li><strong>Ihre Konten, auf Ihren Namen, von Anfang an.</strong> Ein Studio, das Ihre Datenbank auf seinen Namen laufen lässt, hält Ihr Produkt als Pfand – freundlich formuliert.</li>
</ul>
<h2>Häufige Fragen</h2>
<p><strong>Was kostet eine Buchungs-App?</strong> Der funktionierende Kern (Kalender, freie Plätze, Zahlung, Benachrichtigungen, Admin) beginnt bei mir bei etwa 1.000 bis 1.500 Euro. Das Angebot ist kostenlos und kommt mit schriftlichem Umfang.</p>
<p><strong>Ist es riskant, mit einem Entwickler in Korea zu arbeiten?</strong> Das Risiko liegt nicht am Land, sondern am Vertrag. Fragen Sie nach Handelsregisternummer, offizieller Rechnung, Kontoinhaberschaft und einer schriftlichen Ziellinie. Mit diesen vier Punkten spielt die Entfernung keine Rolle.</p>
<p><strong>In welcher Sprache arbeiten wir?</strong> Schriftlich, in Ihrer Sprache. E-Mail und Chat, keine Pflicht-Calls: Seoul und Deutschland trennen acht Stunden, ein Termin verlängert alles nur. Auf Deutsch zu schreiben ist völlig in Ordnung.</p>
<p><strong>Und wenn mir das Ergebnis nicht gefällt?</strong> Korrekturen sind inklusive, bis es passt, und der Restbetrag wird erst nach funktionierender App fällig.</p>
<p><strong>Können Sie im App Store und bei Google Play veröffentlichen?</strong> Ja. Aber die Store-Prüfung ist kein Entwicklungstag: Die App kann an Tag 5 bereit sein, die Freigabe läuft nach der Uhr der Stores.
Sie haben eine Idee, einen hängenden Prototyp oder ein Angebot, das nicht passt? Schreiben Sie mir, und ich sage Ihnen vor jeder Zahlung, um welche Art von Projekt es sich handelt:</p>
]]></content:encoded></item><item><title><![CDATA[Seu app está 80% pronto e travou: o que realmente falta (e quanto custa terminar)]]></title><description><![CDATA[Existe um momento específico em que um projeto gerado com IA para de avançar, e ele quase sempre acontece na mesma tela: o botão de pagamento. O app abre…]]></description><link>https://cpcai.hashnode.dev/seu-app-est-80-pronto-e-travou-o-que-realmente-falta-e-quanto-custa-terminar</link><guid isPermaLink="true">https://cpcai.hashnode.dev/seu-app-est-80-pronto-e-travou-o-que-realmente-falta-e-quanto-custa-terminar</guid><category><![CDATA[startup]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:16:43 GMT</pubDate><content:encoded><![CDATA[<p>Existe um momento específico em que um projeto gerado com IA para de avançar, e ele quase sempre acontece na mesma tela: o botão de pagamento.</p>
<p>O app abre. O login funciona. O calendário mostra os horários. A pessoa clica em «pagar» e recebe um erro 401, ou pior, é redirecionada para uma tela de sucesso que qualquer um pode abrir digitando a URL — e o produto sai de graça.</p>
<p>A partir daí, o dono do projeto faz o que parece lógico: volta para o Lovable, para o Cursor, para o ChatGPT, escreve prompts cada vez mais longos, gera versões novas, e o projeto fica três semanas girando em torno do mesmo problema. Não é falta de ferramenta. é que o que falta não é código novo, é fechar as costuras do código que já existe.</p>
<p>Trabalho sozinho num estúdio de uma pessoa em Seul, na Coreia do Sul, e a maior parte do que recebo é exatamente isso: um app em 80% que precisa ser terminado e publicado. Em 3 a 5 dias, com até 1/20 do preço de um orçamento de agência típico para o mesmo escopo.</p>
<h2>O que está faltando, em ordem</h2>
<p>Quando alguém me manda um projeto travado, eu não começo pelo design. Começo por uma lista de verificação que quase nunca está completa:</p>
<p><strong>1. Autorização, não autenticação.</strong> O app tem login. Quase sempre não tem autorização. Isso significa: eu troco um número na URL e leio o pedido de outra pessoa. Toda IA que gera backend comete esse erro, porque ela testa com uma conta só.</p>
<p><strong>2. O webhook de pagamento.</strong> A confirmação do pagamento tem que acontecer no servidor, não na tela de sucesso. E tem que ser idempotente: se o provedor mandar o mesmo evento duas vezes, o cliente não pode ser cobrado duas vezes nem ganhar dois acessos.</p>
<p><strong>3. Os estados vazios e os estados de erro.</strong> A tela bonita é a tela com dados. O que quebra na vida real é a lista vazia, o nome com acento, o teclado do celular cobrindo o botão de enviar, o botão «voltar» dentro do webview.
<strong>4. O ambiente de produção.</strong> Segredo dentro do repositório, build que só funciona na máquina de alguém, variável de ambiente que ninguém documentou. «Funciona aqui» é metade das vezes um problema de lockfile.</p>
<p><strong>5. A entrega.</strong> Quem é o dono de cada conta, onde cada chave é gerada, como se faz o deploy, como se volta atrás, e qual é a página mais útil de todas: a que lista o que o produto deliberadamente não faz.</p>
<h2>Por que 3 a 5 dias é realista</h2>
<p>Porque o trabalho de um time não está sendo feito. Não tem comercial, gerente de projeto, agência de design e subcontratado entre você e quem escreve o código. Meu próprio site compara com orçamentos típicos de agência e chega a 1/20 do preço; a leitura honesta disso é «orçamento de agência tem muita gente que não constrói», não «app em cinco dias é barato de produzir».
Dia por dia, num projeto de cliente:</p>
<ul>
<li><strong>Dia 1:</strong> um fluxo completo, com dados reais, num link que você abre no celular.</li>
<li><strong>Dia 2:</strong> o dinheiro se move, uma vez, corretamente — ordem, confirmação no servidor, webhook, reembolso.</li>
<li><strong>Dia 3:</strong> duas contas, dois aparelhos, e a conta A tentando ler o registro da B.</li>
<li><strong>Dia 4:</strong> roda no seu domínio, para alguém que não é você, com backup que já foi restaurado uma vez.</li>
<li><strong>Dia 5:</strong> você usa, eu corrijo o que você encontrar, e entrego o documento de handover.</li>
</ul>
<p>O que fica de fora vai por escrito antes de começar: avaliações, cupons, multi-idioma, campanhas agendadas, migração de planilha. Se o orçamento não tem lista de exclusões, os cinco dias não são um escopo — são uma esperança.</p>
<h2>O detalhe que decide entre «barato» e «caro de verdade»</h2>
<p>Um estúdio barato deixa de ser barato no dia em que o financeiro pede uma nota fiscal e não existe nenhuma. Já vi projeto bom morrer nessa etapa, não por causa do código, mas porque o fornecedor era um perfil anônimo sem empresa por trás.
Minhas condições são verificáveis antes de você pagar qualquer coisa:</p>
<ul>
<li><strong>Zero antes do contrato.</strong> Revisar o projeto e estimar é gratuito. Se a resposta for «isso precisa de três semanas e eu explico por quê», você perdeu uma hora, não um sinal.</li>
<li><strong>Resto depois de ver o app rodando.</strong> Sinal na assinatura, o restante quando os dias acima já aconteceram.</li>
<li><strong>Contrato com empresa coreana de verdade e nota fiscal.</strong> CPC AI Inc., registro 771-88-03539, em Seul.</li>
<li><strong>Suas contas, no seu nome, desde o começo.</strong> Estúdio que guarda seu banco de dados no nome dele está segurando seu produto como refém, gentilmente.</li>
</ul>
<h2>Perguntas que sempre recebo</h2>
<p><strong>Quanto custa terminar um app que já está em 80%?</strong> Depende de quantas costuras estão abertas, e a revisão é gratuita justamente para isso. O núcleo funcional de um app de reservas com pagamento começa na faixa de 1.000–1.500 dólares, com escopo e preço fechados por escrito antes do contrato.</p>
<p><strong>Funciona para quem está no Brasil?</strong> Sim. Todo o trabalho é por escrito, em português, por e-mail e chat. A diferença de horário entre Seul e São Paulo é de 12 horas, então marcar reunião é pior do que escrever: normalmente a resposta chega no mesmo dia útil seu.</p>
<p><strong>E se eu não gostar do resultado?</strong> As revisões estão incluídas até ficar como você precisa, e o pagamento final só acontece depois que você vê o app funcionando.</p>
<h2><strong>Vocês publicam na App Store e no Google Play?</strong> Sim. Só entenda que a análise da loja não é um dia de programação: o app fica pronto para publicar no dia cinco, mas a aprovação anda no relógio da loja.</h2>
<p>Se você tem um app parado em 80%, uma ideia que precisa sair do papel ou um orçamento que não fechou, escreva e eu digo em qual tipo de trabalho você está antes de qualquer pagamento:</p>
<p><strong><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></strong> — revisão gratuita, escopo e preço fechados depois.
<strong><a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></strong></p>
<p>CPC AI Inc. · Seul, Coreia do Sul · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[Cuánto cuesta desarrollar una app en 2026: la parte de la factura que nadie te muestra]]></title><description><![CDATA[Casi todas las semanas me llega el mismo mensaje, con distintas palabras: «tengo la idea, tengo el diseño en Figma y tengo un presupuesto de 12.000…]]></description><link>https://cpcai.hashnode.dev/cuanto-cuesta-desarrollar-una-app-2026</link><guid isPermaLink="true">https://cpcai.hashnode.dev/cuanto-cuesta-desarrollar-una-app-2026</guid><category><![CDATA[startup]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:14:02 GMT</pubDate><content:encoded><![CDATA[<p>Casi todas las semanas me llega el mismo mensaje, con distintas palabras: «tengo la idea, tengo el diseño en Figma y tengo un presupuesto de 12.000 dólares. ¿Por qué dos agencias me pidieron 40.000 y tres meses?»</p>
<p>La pregunta parece de precio. En realidad es una pregunta sobre cuántas personas hay entre tú y la persona que escribe el código.</p>
<p>Trabajo solo. Soy el único desarrollador de CPC AI Inc., un estudio de una persona en Seúl, y construyo aplicaciones web y móviles con IA en 3 a 5 días. No tengo comerciales, ni jefe de proyecto, ni agencia de diseño, ni subcontratistas. Eso es lo que explica la diferencia de precio, y también lo que explica por qué el plazo se mide en días y no en trimestres.</p>
<h2>El número que todos preguntan primero</h2>
<p>Para la misma app de reservas con pagos (calendario, cupos, pago con tarjeta, avisos por correo, panel de administración):</p>
<ul>
<li><strong>Agencia tradicional:</strong> entre 15.000 y 40.000 dólares, 3 a 6 meses, equipo de 4 a 8 personas.</li>
<li><strong>Freelance senior en LatAm o España:</strong> entre 4.000 y 12.000 dólares, 4 a 10 semanas, una persona pero con agenda compartida.</li>
<li><strong>Un estudio de una persona que construye directamente:</strong> desde unos 1.000–1.500 dólares y 3 a 5 días para el núcleo funcional.
Mi propio sitio compara presupuestos típicos de agencia para el mismo alcance y lo pone en «hasta 1/20 del precio». La lectura honesta de ese número no es «es barato producir una app en cinco días». Es «en un presupuesto de agencia hay mucha gente que no está construyendo». Cuando quitas esas capas, lo que queda es tiempo real de ingeniería, y el tiempo real de ingeniería de una app de reservas es de días, no de meses.</li>
</ul>
<h2>Lo que sí cambia el precio (y no es la calidad)</h2>
<p>Después de varios proyectos encontré tres cosas que mueven el número mucho más que la tecnología:</p>
<p><strong>1. Cuántas decisiones tienes que tomar tú.</strong> Un proyecto donde el cliente contesta en dos horas avanza cinco veces más rápido que uno donde la pregunta «¿el precio es por clase o por persona?» queda dos días sin respuesta. Ese silencio no aparece en la cotización, pero aparece en el calendario.</p>
<p><strong>2. Si las cuentas están a tu nombre.</strong> Dominio, base de datos, pasarela de pago, ficha en la tienda de aplicaciones. Si están en nombre del estudio, arreglarlo es el punto uno del trabajo, no un detalle.</p>
<p><strong>3. Qué tan lejos está de «ya casi».</strong> Una app hecha con Lovable o Cursor al 80% no es un 20% de trabajo. Es un trabajo distinto: revisar permisos, verificar que un webhook repetido no cobre dos veces, y separar el entorno de producción del de pruebas.</p>
<h2>Qué es exactamente un MVP de 3 a 5 días</h2>
<p>No es una demo. Es una app que puedes abrir en tu teléfono y usar con datos reales. Día por día, en proyectos de cliente:</p>
<p><strong>Día 1 — un flujo completo, con datos reales.</strong> No la app entera: registro, la pantalla principal y una escritura que sobrevive a recargar la página. Se ve sencilla a propósito; el diseño bonito el día uno cuesta un día, el día cinco cuesta una hora.</p>
<p><strong>Día 2 — el dinero se mueve, una vez, correctamente.</strong> Creación de la orden, confirmación del pago en el servidor, webhook y ruta de reembolso. La prueba al final del día: un pago real da exactamente un acceso, una vez. Si el webhook llega dos veces, el cliente no paga dos veces.
<strong>Día 3 — las costuras que aguantan un segundo usuario.</strong> Dos cuentas: la cuenta A intenta leer el registro de B cambiando un número en la URL. Si devuelve datos, todo el backend necesita una pasada de autorización.</p>
<p><strong>Día 4 — corre en tu dominio, para alguien que no eres tú.</strong> Variables de entorno fuera del repositorio, despliegue desde cero, certificado, monitoreo de errores, copia de seguridad que alguien ya restauró una vez.</p>
<p><strong>Día 5 — lo usas tú, yo arreglo lo que encuentres.</strong> Aquí aparecen los dos hechos que ningún documento recogió: tu equipo siempre reserva desde una tablet, o los nombres de tus clases llevan una barra que rompe la URL.</p>
<p>Y lo que <strong>no</strong> entra, por escrito y antes de empezar: reseñas, cupones, multi-idioma, campañas programadas, migración de datos desde Excel. Si una cotización no tiene lista de exclusiones, los cinco días no son un alcance: son un deseo.</p>
<h2>Si tu app ya está al 80%</h2>
<p>Es el caso más común que recibo: un prototipo generado con IA que funciona en local, el botón de pago devuelve 401, y el proyecto lleva tres semanas parado. En ese punto, escribir más prompts no arregla nada. Lo que falta es terminar las costuras en orden y decirte con honestidad qué tipo de trabajo tienes entre manos.</p>
<h2>El contrato y la factura (la parte aburrida que decide todo)</h2>
<p>Un estudio barato deja de ser barato en el momento en que tu contabilidad pide un documento y no hay ninguno. Mis condiciones son verificables antes de pagar nada:</p>
<ul>
<li><strong>0 dólares antes del contrato.</strong> Revisar el proyecto, estimar y planear es gratis. Si mi respuesta es «esto necesita tres semanas y te explico por qué», perdiste una hora, no un depósito.</li>
<li><strong>Saldo después de ver software funcionando.</strong> Depósito al firmar, el resto cuando los días anteriores ya pasaron y ya hiciste clic.</li>
<li><strong>Contrato con una empresa coreana real y factura oficial.</strong> CPC AI Inc., número de registro de empresa 771-88-03539, en Seúl. No es un perfil anónimo de una plataforma de freelancers.</li>
<li><strong>Tus cuentas, a tu nombre, desde el principio.</strong></li>
</ul>
<h2>Preguntas frecuentes</h2>
<p><strong>¿Cuánto cuesta una app de reservas o de pedidos?</strong> El núcleo funcional (calendario, cupos, pago, notificaciones, panel) arranca alrededor de 1.000–1.500 dólares en mi caso, y depende sobre todo de cuántos flujos distintos haya y de si hay que migrar datos. La cotización es gratis y llega con un alcance escrito.</p>
<p><strong>¿Es seguro contratar a un desarrollador en Corea?</strong> La seguridad no viene del país, viene del contrato. Pide número de registro de empresa, factura oficial, titularidad de las cuentas y una línea de meta escrita. Con esas cuatro cosas, la distancia es irrelevante.</p>
<p><strong>¿En qué idioma trabajamos?</strong> Por escrito, en tu idioma. Correo y chat, sin llamadas obligatorias: la diferencia horaria entre Seúl y América es de 12 a 14 horas y las llamadas solo alargan las cosas. Escribir en español está perfectamente bien.
<strong>¿Qué pasa si no me gusta el resultado?</strong> Las revisiones están incluidas hasta que quede como lo necesitas, y el saldo se paga después de ver la app funcionando. Ese orden existe para que el riesgo no sea todo tuyo.</p>
<p><strong>¿Puedo lanzar a App Store y Google Play?</strong> Sí, pero la revisión de la tienda no es un día de programación: la app puede estar lista para publicar el día cinco y la ficha sale en el reloj de la tienda, no en el mío.</p>
<hr />
<p>Si tienes una idea, un prototipo atascado o un presupuesto que se quedó corto, escríbeme y te digo en qué clase de trabajo estás antes de que pagues nada:</p>
<p><strong><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></strong> — revisión gratuita, alcance y precio fijo después.
<strong><a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></strong></p>
<p>CPC AI Inc. · Seúl, Corea del Sur · 771-88-03539</p>
]]></content:encoded></item><item><title><![CDATA[Four payment rails that decide your app's architecture before you write a line]]></title><description><![CDATA[Most app briefs I get from outside the US and Western Europe mention payments in one line: "integrate card payment". Then the build starts, and the payment…]]></description><link>https://cpcai.hashnode.dev/four-payment-rails-that-decide-your-app-s-architecture-before-you-write-a-line</link><guid isPermaLink="true">https://cpcai.hashnode.dev/four-payment-rails-that-decide-your-app-s-architecture-before-you-write-a-line</guid><category><![CDATA[payments]]></category><category><![CDATA[mobile app development]]></category><category><![CDATA[fintech]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 18:02:26 GMT</pubDate><content:encoded><![CDATA[<p>Most app briefs I get from outside the US and Western Europe mention payments in one line: "integrate card payment". Then the build starts, and the payment rail turns out to be the part that decides what the app can even do. Not the design, not the framework. The rail.</p>
<p>I build apps alone from Seoul, at CPC AI Inc. (Korean company registration 771-88-03539), and I have shipped two of my own: the mobile strategy game Cannon King and the crypto market dashboard CPCAI. Before that I spent years watching small businesses buy software that worked in a demo and fell apart at the payment step. So here are the four rails I now design for explicitly, and what each one forces into the architecture.</p>
<h2>1. UPI (India): a payment is an intent, not a card charge</h2>
<p>UPI is not "a wallet". It is an account-to-account request/collect protocol with a lifecycle: intent created, approved or declined by the user in their own app, then a status that has to be reconciled server-side. The moment a client says UPI, three things change:</p>
<ul>
<li>The app needs a server callback endpoint that verifies the payment signature, because the client-side "success" screen is not proof.</li>
<li>Timeouts are normal. Users abandon the UPI app, come back, retry. Your order table needs a state machine, not a boolean.</li>
<li>Refunds and partial captures are handled by the PSP, not by you, and they arrive asynchronously.</li>
</ul>
<p>If you build UPI as "call API, show checkmark", you will ship an app that gives away the product for free during the first network hiccup. I wrote about this in a Hindi piece for Indian founders last week, because it is the single most common defect I find in "finished" AI-built Indian apps.</p>
<h2>2. Pix (Brazil): instant, irrevocable, and it changed checkout design</h2>
<p>Pix settlements are near-instant and, in practice, irreversible. That sounds like a gift until you notice what it does to the return flow: there is no chargeback to lean on, so the app has to be stricter before it grants access. The pattern that works is a short-lived QR with a server-verified webhook, plus an explicit "expired, generate a new one" state that users actually see. Brazilian users expect the QR to be scanned from a second screen, so a full-screen QR with a countdown beats a modal dialog.
The other thing Pix changed is pricing psychology: because the payment is instant and free for consumers, a "test it for a week then pay" model feels slow. In Brazil, the honest version is "pay a small deposit, see it working, pay the rest" — the same model I use everywhere, but the messaging has to match a market that moves in seconds.</p>
<h2>3. Raast (Pakistan) and M-Pesa (Kenya): the rail is domestic, the client is not</h2>
<p>Here is the structural problem in Pakistan, Kenya, Nigeria, Bangladesh and a dozen other markets: local rails are excellent and cheap, but a foreign studio cannot be paid through them. Raast moves rupees between Pakistani accounts. M-Pesa moves shillings between Safaricom wallets. Neither reaches a Korean company's account.</p>
<p>So the build has two payment systems, not one:</p>
<ul>
<li>The customer-facing rail inside the app (Raast, JazzCash, Easypaisa, M-Pesa), which is fast, local and cheap.</li>
<li>The business-to-studio rail outside the app (wire transfer, invoice in USD), which is slow, documented and needs a paper trail.
Designing only the first one is why so many projects stall after the demo. The app works, the client loves it, and then nobody can move money across the border without three weeks of bank questions. I now ask for the invoice requirements on day one, and I put the amount in USD in the contract so the exchange rate does not reopen the price later.</li>
</ul>
<h2>4. MilliQR, LankaQR, Bakong, QPay, IPS QR: QR standards that are not interchangeable</h2>
<p>Every one of these is a national QR standard with its own payload format, its own merchant onboarding and its own reconciliation file. They look identical in a screenshot — a square that a phone scans — and share almost nothing underneath.</p>
<p>The practical consequence: "add a QR payment" is not a five-hour task in Baku, Colombo, Phnom Penh or Ulaanbaatar. It is a task that starts with the client's bank, because the merchant account has to exist before any SDK is useful. I have lost days to this, and now the first deliverable of any QR project is a screenshot of the merchant account in test mode, sent by the client, before I write the checkout screen.</p>
<h2>What this does to a five-day build</h2>
<p>None of this makes a 3–5 day build impossible. It changes what day one is for. Day one is not screens; day one is four answers:</p>
<ol>
<li>Which rail actually moves money for this business, and is the merchant account already approved?</li>
<li>Who owns the App Store and Google Play account, and can they pay the $99 / $25 fees from that country?</li>
<li>What document does the client's accountant need from a foreign vendor, and does it need to be issued in USD or local currency?</li>
<li>What happens on the device the customer actually owns — mid-range Android, unstable data, app size under 30 MB?</li>
</ol>
<p>Get those four and the rest is ordinary work: one main flow (choose, pay, confirm), an admin panel, the store's required pages, and a handover where the accounts and the database end up in the client's name, not mine.</p>
<p>The reason a solo studio can do this in days instead of months is not that the work is small. It is that there is no sales team, no project manager and no handoff between three departments — the same person who asks the payment questions writes the code that answers them.</p>
<p>If you are building for a market where the rail is not Stripe, and the last 20% is stuck at the payment step, that is exactly the part I take over. Site: <a href="https://cpcai-studio.netlify.app">https://cpcai-studio.netlify.app</a> — email: <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a>.</p>
]]></content:encoded></item><item><title><![CDATA[The invoice is where a cheap overseas studio usually stops being cheap]]></title><description><![CDATA[There is a moment in every cross-border project that nobody puts on the timeline. The app works, the client is happy, and then someone in finance asks for…]]></description><link>https://cpcai.hashnode.dev/the-invoice-is-where-a-cheap-overseas-studio-usually-stops-being-cheap</link><guid isPermaLink="true">https://cpcai.hashnode.dev/the-invoice-is-where-a-cheap-overseas-studio-usually-stops-being-cheap</guid><category><![CDATA[Freelancing]]></category><category><![CDATA[startup]]></category><category><![CDATA[business]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 06:02:59 GMT</pubDate><content:encoded><![CDATA[<p>There is a moment in every cross-border project that nobody puts on the timeline. The app works, the client is happy, and then someone in finance asks for a document the studio has never issued in English, and the last payment sits in a queue for two weeks while a bank in one country tries to work out what a Korean business registration number is.</p>
<p>I have watched that happen from the studio side. It is boring, it is avoidable, and it is a better test of whether you are dealing with a real company than any portfolio.</p>
<h2>1. Ask what the invoice will say before you agree the price</h2>
<p>A one-person studio that is not a company can build you an excellent app. It cannot easily give you an invoice with a corporate registration number on it, and that number is what your accountant needs to book the payment as a service rather than a transfer to a stranger.</p>
<p>The document should carry four things, and you can ask for a sample on day one:</p>
<ul>
<li>the legal entity name and its registration number (ours is <strong>CPC AI Inc., 771-88-03539</strong>),</li>
<li>the address the entity is registered at,</li>
<li>a description of the service specific enough to survive an audit ("mobile app development, 5 days, scope attached" rather than "consulting"),</li>
<li>the currency, the amount, and the payment terms written the same way they were agreed in chat.</li>
</ul>
<p>If a studio cannot produce that within a day, the price you were quoted is not the price you will end up explaining to your finance team.</p>
<h2>2. Currency: the number in the chat and the number in the bank are not always the same</h2>
<p>Three things move between the quote and the money leaving your account: your bank's conversion spread, the intermediary bank's fee, and whatever your card issuer adds if you pay by card instead of wire.</p>
<p>This is not a reason to avoid overseas studios — it is a reason to agree one number, in one currency, with the fees assigned to someone. In practice:</p>
<ul>
<li><strong>Wire</strong> is cheapest for a large balance and slowest to clear. Ask who pays the sending fee; assume the receiving side eats its own.</li>
<li><strong>Card</strong> is fastest and adds a processing percentage. If the studio accepts card, that percentage is usually already inside the price, which is fine as long as you know it.</li>
<li><strong>Anything that requires the studio to hold your money</strong> — escrow, holding accounts — either does not exist for a five-day build or costs more than the build's first day.
For a build this small, the useful rule is simpler than any of that: <strong>the amount at risk at each step should be small enough that a disagreement does not need a lawyer</strong>. That is why the studio I run takes nothing before a contract, shows the running app on your own accounts, and asks for the balance after you have clicked through it yourself.</li>
</ul>
<h2>3. Tax forms are a real step, not paperwork theatre</h2>
<p>If you are outside Korea, your finance team may ask for a tax residency form before they release a payment to a foreign vendor — a W-8BEN or a W-8BEN-E if you are in the United States, an equivalent residence certificate in many other places. Studios that have done cross-border work have these ready and can fill them in an afternoon. Studios that have not will discover the requirement on the day the payment is due, and you will be the one chasing it.</p>
<p>Ask in the first call: "have you been paid from outside Korea before, and can you sign a tax form for our finance team?" The answer tells you more about the studio's maturity than the case studies do.</p>
<h2>4. The handover and the money should point at each other</h2>
<p>The last payment is where overseas projects go wrong, and almost always because the two things being exchanged were never defined precisely:</p>
<ul>
<li>what physically moves (source code in a repository you own, the database and hosting under your billing account, the domain in your name, store seller accounts),</li>
<li>when it moves (at handover, not at "launch"),</li>
<li>and what happens to the money if what moves is not what was agreed.</li>
</ul>
<p>Write it as three lines in the scope. It takes five minutes and it is the reason the final week is calm rather than tense.</p>
<h2>5. What a short build changes about all of this</h2>
<p>A three-to-six month project cannot be paid in steps small enough to feel safe, so both sides end up trusting a contract and a lawyer. That is the honest reason long projects demand a legal frame.</p>
<p>A three-to-five day build is different: the whole thing is one milestone, the amount is at most a twentieth of the long quote, and the evidence arrives before the balance instead of after it. That is not a payment trick, it is what makes an overseas studio hireable by a small company at all — and it is why I ask for the awkward documents on day one rather than day five.</p>
<p>If you are comparing an overseas quote with a local one and the numbers look too far apart to be the same product, the invoice is a good place to start the comparison: ask both sides what their document will say.</p>
<ul>
<li>Studio (English): <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></li>
<li>Studio (Japanese): <a href="https://cpcai-studio.netlify.app/ja/">https://cpcai-studio.netlify.app/ja/</a></li>
<li>Email, and you get a reply in your own language: <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></li>
<li>CPC AI Inc. (Seoul, Korea) — business registration no. 771-88-03539</li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Row-level security that looks fine and leaks: three Supabase patterns AI code generators leave behind]]></title><description><![CDATA[If an AI tool built your Supabase backend, there is a specific kind of bug waiting for you, and it will not show up in testing with one account. The…]]></description><link>https://cpcai.hashnode.dev/row-level-security-that-looks-fine-and-leaks-three-supabase-patterns-ai-code-generators-leave-behind</link><guid isPermaLink="true">https://cpcai.hashnode.dev/row-level-security-that-looks-fine-and-leaks-three-supabase-patterns-ai-code-generators-leave-behind</guid><category><![CDATA[supabase]]></category><category><![CDATA[webdev]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Mon, 16 Nov 2026 01:48:06 GMT</pubDate><content:encoded><![CDATA[<p>If an AI tool built your Supabase backend, there is a specific kind of bug waiting for you, and it will not show up in testing with one account. The policies exist. They are enabled. The app behaves perfectly for the person who built it, because that person is the only user, and they own every row.</p>
<p>Here are the three shapes I keep finding when I take over an "almost finished" app, and the twenty-minute check that catches all of them.</p>
<h2>1. <code>USING (true)</code> for authenticated users</h2>
<p>This is the policy most templates generate, because it makes the demo work:</p>
<pre><code class="language-sql">alter table public.tasks enable row level security;

create policy "Enable read access for all users"
on public.tasks for select
to authenticated
using (true);
</code></pre>
<p>It reads like "logged-in users can read their tasks". It actually means "any logged-in user can read every row in the table". With one test account you will never see it: your tasks are the only tasks.</p>
<p>The check is mechanical: compare the tables that hold user data against the policies that mention <code>auth.uid()</code>.</p>
<pre><code class="language-sql">select tablename, policyname, cmd, qual, with_check
from pg_policies
where schemaname = 'public'
order by tablename;
</code></pre>
<p>If <code>qual</code> is null or <code>true</code> on a table with a <code>user_id</code> column, that table is shared. On a hobby project it is embarrassing. On anything with names, addresses, invoices, or uploads it is a reportable problem, and in the EU it is the kind of thing that has a deadline attached.</p>
<p>The fix is one policy per verb, not per table:</p>
<pre><code class="language-sql">create policy "own rows: select" on public.tasks for select
to authenticated using (user_id = auth.uid());

create policy "own rows: insert" on public.tasks for insert
to authenticated with check (user_id = auth.uid());

create policy "own rows: update" on public.tasks for update
to authenticated using (user_id = auth.uid()) with check (user_id = auth.uid());

create policy "own rows: delete" on public.tasks for delete
to authenticated using (user_id = auth.uid());
</code></pre>
<p><code>with check</code> is the half people skip. Without it an update policy lets a row be read and then rewritten with someone else's <code>user_id</code>, which is a quiet way to move data between accounts.</p>
<h2>2. The view or RPC that runs as the owner</h2>
<p>AI code will happily create a view for a dashboard, and a view in Postgres runs with the privileges of its owner unless you say otherwise. RLS on the base table does not apply, and the dashboard politely returns all rows.</p>
<p>The same trap appears as a function:</p>
<pre><code class="language-sql">create or replace function public.tasks_for_user(uid uuid)
returns setof tasks
language sql security definer
as $$ select * from tasks where user_id = uid $$;
</code></pre>
<p>Three problems in six lines: <code>security definer</code>, a caller-supplied <code>uid</code>, and no <code>search_path</code>. Any authenticated user can call <code>rpc('tasks_for_user', { uid: '&lt;someone else&gt;' })</code> and get that person's rows. If the function were <code>security invoker</code> the RLS policies would apply and the <code>uid</code> argument would be pointless — which is the point of RLS: the query does not need to be trusted, the table is.</p>
<pre><code class="language-sql">-- if you must have a definer function, pin the identity inside it
create or replace function public.my_tasks()
returns setof tasks
language sql security definer set search_path = public
as $$ select * from tasks where user_id = auth.uid() $$;
</code></pre>
<p>And revoke execute from <code>anon</code> on anything you do not want called before sign-in.</p>
<h2>3. The key in the wrong place</h2>
<p>Every Supabase project has an <code>anon</code> key and a <code>service_role</code> key. The <code>anon</code> key is meant to be public. The <code>service_role</code> key bypasses RLS completely, and it is the single most expensive string to leak.</p>
<p>It leaks in a shape that looks responsible: someone adds a "secure" API route to keep the key off the client, and the route forwards whatever the client sends:</p>
<pre><code class="language-ts">// looks like a backend; is a proxy for anybody's requests
app.post('/api/delete-task', async (req) =&gt; {
  const admin = createClient(url, process.env.SUPABASE_SERVICE_ROLE_KEY);
  return admin.from('tasks').delete().eq('id', req.body.id);
});
</code></pre>
<p>No auth check, no ownership check, and the key is now reachable through an endpoint that anyone can call with any id. A <code>service_role</code> call should only exist where the server has already proved who the caller is — a verified session, a webhook signature, or a cron job.</p>
<h2>The twenty-minute audit</h2>
<p>I run this in this order on every takeover, before touching any feature:</p>
<ol>
<li><strong>List what has RLS off.</strong> <code>select tablename from pg_tables where schemaname = 'public' and rowsecurity = false;</code> Anything with user data here is wide open to the <code>anon</code> key.</li>
<li><strong>List policies whose <code>qual</code> does not mention <code>auth.uid()</code></strong> on tables that have a <code>user_id</code> or <code>owner_id</code> column.</li>
<li><strong>Find every <code>security definer</code> function</strong> and check whether it takes an identity from the caller: <code>select proname, prosrc from pg_proc where prosecdef;</code></li>
<li><strong>Test as a second person, not as the owner.</strong> In the SQL editor you can imitate a signed-in user without their password:</li>
</ol>
<pre><code class="language-sql">set local role authenticated;
set local request.jwt.claims = '{"sub":"&lt;other-user-uuid&gt;","role":"authenticated"}';
select count(*) from public.tasks;  -- should be 0, not the owner's row count
</code></pre>
<p>If that last count is not zero, nothing else on the list matters yet.</p>
<ol start="5">
<li><strong>Check the buckets.</strong> A public storage bucket with guessable names is the same leak with files, and AI tools default to public when the upload needs to show up in an <code>&lt;img&gt;</code> tag. Private bucket plus signed URLs is two lines of code and the difference between a leak and a nuisance.</li>
</ol>
<h2>Why this is the expensive 20%</h2>
<p>None of this is hard. It is two or three days of reading code you did not write, deciding what "correct" means for each table, and testing with a second account, a second browser, and a phone.</p>
<p>It is also the part that never appears on a feature list, because nothing on the screen changes when you fix it. What changes is that you can hand the link to someone who is not you — and that is the only test that matters before you charge money for an app.</p>
<p>I do this as a one-person studio in Seoul: I take an app that already looks finished, keep what works, and get it to the point where a stranger can sign up, pay, and come back tomorrow — three to five days for a focused web or mobile app, with the review of the data model and policies included rather than sold separately.</p>
<ul>
<li>Studio (English): <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></li>
<li>Japanese: <a href="https://cpcai-studio.netlify.app/ja/">https://cpcai-studio.netlify.app/ja/</a></li>
<li>Email: <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></li>
</ul>
<p>If you want a starting point rather than a read: open your Supabase project, run the four queries above, and look at the first table that comes back with <code>rowsecurity = false</code>. That is your next two days, and it is a better use of them than the feature you were about to add.</p>
]]></content:encoded></item><item><title><![CDATA[AI MVP development in 3–5 days: what actually lands in your hands each day]]></title><description><![CDATA[Everyone who writes to me asks the same question: how fast can you finish it? Almost nobody asks the better one: what do I receive on Tuesday? That…]]></description><link>https://cpcai.hashnode.dev/ai-mvp-development-in-3-5-days-what-actually-lands-in-your-hands-each-day</link><guid isPermaLink="true">https://cpcai.hashnode.dev/ai-mvp-development-in-3-5-days-what-actually-lands-in-your-hands-each-day</guid><category><![CDATA[AI]]></category><category><![CDATA[startup]]></category><category><![CDATA[webdev]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sun, 15 Nov 2026 16:22:07 GMT</pubDate><content:encoded><![CDATA[<p>Everyone who writes to me asks the same question: how fast can you finish it?</p>
<p>Almost nobody asks the better one: what do I receive on Tuesday?</p>
<p>That question decides whether a 3–5 day AI MVP development timeline is real or a sales line. A speed claim tells you nothing until you can name what appears on screen on day two, who has to do something for it to appear, and what is still missing when the five days are over. So here is the day-by-day version, as it runs on client projects, with the parts that go wrong written in.</p>
<p>Context first, because it sets the scale: I am one person. I build with Claude, Cursor-era tooling and a stack I have shipped before, and I have run two products of my own to store release, so a client project is not my first deploy. There is no agency behind me, no account manager, and no one to forward your question to. That is why the timeline is days instead of months.</p>
<h2>Day 1 — One flow works end to end, with real data</h2>
<p>Not the whole app. One flow, finished, on a real URL: usually sign-up or login, the core screen, and a write that survives a refresh.
By the end of day 1 you can open the link on your phone and click through something. It will look plain. That is deliberate — design polish on day five costs an hour, design polish on day one costs a day because the screens will change.</p>
<p>This is also the day the questions you did not know you had start arriving. "Wait, the price is per class or per person?" is cheap on day 1 and expensive on day 4.</p>
<p>What I need from you before day 1 starts, and no later:</p>
<ul>
<li><strong>The one user you are building for</strong>, named as a person or a job title, not "our customers".</li>
<li><strong>The accounts you own</strong>: domain, database, payment provider, store listing, hosting. If they are in someone else's name, fixing that is work item one.</li>
<li><strong>A decision-maker who answers within a few hours.</strong> The single biggest cause of a five-day build becoming a three-week build is a question that sits unanswered for two days.</li>
</ul>
<h2>Day 2 — Money moves, once, correctly</h2>
<p>Payments are where AI-built apps usually die. The generated code has a beautiful checkout button and a redirect that grants access to anyone who lands on the success URL — which means anyone can get your product for free by typing a URL.</p>
<p>Day 2 is order creation, payment confirmation on the server, the webhook, and the refund path. The test at the end of the day is simple: one real payment grants exactly one entitlement, once. If a webhook arrives twice, the customer is not charged twice or granted twice.</p>
<p>This is also the day a genuine dependency can stop the clock: merchant approval, business verification, a payment provider that does not serve your country, or a Korean PG whose review takes days. I do not hide that inside the timeline. It goes in the estimate as calendar risk, separate from build days, and we start it before day 1 where possible.</p>
<h2>Day 3 — The seams that hold under a second user</h2>
<p>Two accounts, one device each. Account A tries to read B's record by changing an id in the URL. If it returns data, every endpoint needs an authorization pass, and that is the day's work.</p>
<p>Then the unglamorous seams: notification delivery, double-booking guards, empty states, error states, the phone keyboard covering the submit button, the back button in a web view, a session that survives a refresh and a cold start.
Most "it worked yesterday" reports come from this layer. AI-generated code is usually structurally neat and dishonest about exactly these boundaries, because the generator never ran the app in the environment it would ship to.</p>
<h2>Day 4 — It runs on your domain, for someone who is not you</h2>
<p>Deployment day: environment separation, secrets out of the repo, a clean clone deployed from scratch, the domain and certificate, error monitoring, a backup that has actually been restored once, and a rollback command that is one line long.</p>
<p>"Works on my laptop" is a lockfile problem about half the time. The other half is an environment variable nobody documented and a build setting that lives in someone's browser.</p>
<h2>Day 5 — You use it, I fix what you find</h2>
<p>Handover day is not padding. It is the day you sit with the thing and discover the two facts no specification contained: your staff always book from a tablet, or your class names contain a slash that breaks the URL.</p>
<p>Then the handover itself, in a form you can keep: who owns which account, where each key is issued, how to deploy, how to roll back, what support exists, and — the most useful page — what the product deliberately does not do.</p>
<p>If you are releasing to the App Store or Google Play, the store review is not a coding day and cannot be compressed by me or by anyone. The build can be store-ready on day five; the listing goes live on the store's clock.</p>
<h2>What is not in the five days</h2>
<p>Every honest 3–5 day AI MVP development quote has a subtraction list, and it should be in writing before work starts. Mine usually says: reviews and ratings deferred, coupons deferred, multi-language deferred, scheduled campaigns deferred, and any migration of existing spreadsheet data quoted separately — because moving 400 customers out of Excel is a different job with a deduplication problem in it, not a screen.</p>
<p>If a quote has no subtraction list, the five days are not a scope, they are a hope.</p>
<h2>The cost part</h2>
<p>Here is the reason the schedule can be this short, and it is not magic: the work of a sales team, a PM, a design agency and a subcontractor is not being done. There is no margin stacked between you and the person writing the code. My own site puts the comparison against typical agency quotes for the same scope at up to 1/20 of the price; the honest reading of that number is "typical quotes contain a lot of people who are not building", not "a five-day build is cheap to produce".</p>
<p>The commercial terms matter more than the number, and they should be checkable rather than trusted:</p>
<ul>
<li><strong>Nothing before the contract.</strong> Reviewing the project, the estimate and the plan costs you zero. If the answer is "this needs three weeks and here is why", you have lost an hour, not a deposit.</li>
<li><strong>Balance after you see running software.</strong> A deposit at contract, the rest after the days above have happened and you have clicked through them.</li>
<li><strong>A written finish line.</strong> What "done" means in one sentence, which seams get verified, what counts as a revision and what counts as new work.</li>
<li><strong>Your accounts, in your name, from the start.</strong> A studio that keeps your database in its own name is holding your product hostage, kindly.</li>
</ul>
<h2>How to use this if you are shopping for a finish</h2>
<p>Ask any studio offering 3–5 days three questions, and listen for specifics:</p>
<ol>
<li>What will I see on day 2, and what will still be missing?</li>
<li>Who owns the payment account and the store listing at the end?</li>
<li>What is not included this week, in writing?</li>
</ol>
<p>A real shop answers in minutes with names of things. A reseller answers with adjectives.</p>
<p>If you already have an app sitting at "almost there" — the Lovable build that works locally, the Cursor project stuck in store review, the Bolt prototype whose payment button returns 401 — the fast path is not more prompting. It is somebody finishing the seams above, in order, and telling you honestly which class of job you have.</p>
<p>That is what I do, one project at a time, directly:</p>
<p><strong><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></strong> — free review first, fixed finish line and price after that. One-person studio in Seoul, formal contract, official invoice, no agency in between. Questions in English are fine: <strong><a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></strong></p>
]]></content:encoded></item><item><title><![CDATA[The handover question nobody asks until month four]]></title><description><![CDATA[What handover means: 12 items that decide whether you can run your own product after the developer leaves, plus three contract lines that cover it.]]></description><link>https://cpcai.hashnode.dev/the-handover-question-nobody-asks-until-month-four</link><guid isPermaLink="true">https://cpcai.hashnode.dev/the-handover-question-nobody-asks-until-month-four</guid><category><![CDATA[Startups]]></category><category><![CDATA[webdev]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sun, 15 Nov 2026 05:18:21 GMT</pubDate><content:encoded><![CDATA[<p>A founder called me in month four. His app worked. His customers were paying. He wanted to change the logo on the login screen and could not, because the person who built the app had stopped answering messages and nobody had the password to the server.</p>
<p>Nothing was broken. He owned a business he could not operate.</p>
<p>This is the most common way a software project goes wrong after launch, and it is almost never discussed while the project is being bought. Buyers ask about price, timeline, and whether you have done something like this before. Nobody asks who will be able to deploy at 9 a.m. on a Monday, six months from now.</p>
<p>Here is the checklist I would use if I were hiring someone, written by someone who gets hired.</p>
<h2>Handover is not a zip file</h2>
<p>Code you cannot run is a folder. What actually decides whether you can operate your product is smaller and more boring than the code:</p>
<ol>
<li><strong>The repository</strong>, in your account, not the vendor's. If the vendor disappears, the code and its history stay.</li>
<li><strong>The deploy target</strong> — Vercel, Netlify, a VPS — on your billing account.</li>
<li><strong>The domain and DNS</strong>, with the registrar login in your hands and the records written down.</li>
<li><strong>The database</strong>, on your instance, with credentials you hold and an export you can open.</li>
<li><strong>File storage</strong> — images, uploads, documents — in your bucket.</li>
<li><strong>The payment account</strong> under your legal entity, so the money never passes through someone else's dashboard.</li>
<li><strong>Email and SMS providers</strong>: your API keys, your quota, your sender domain.</li>
<li><strong>Third-party services</strong> — maps, push, analytics, AI APIs — listed in one place with owner and renewal date.</li>
<li><strong>Secrets</strong> in your secret store, with a note on which key does what.</li>
<li><strong>Admin access</strong>: how to sign in as an administrator and where user roles are defined.</li>
<li><strong>Backups</strong>: a schedule, and one restore actually performed. An untested backup is a wish.</li>
<li><strong>A one-page runbook</strong>: how to deploy, how to roll back, who to call.</li>
</ol>
<p>Twelve items. In my experience most post-launch disputes are not about code quality. They are about item two and item six.</p>
<h2>Three contract lines that prevent most of it</h2>
<p>Korean contracts with freelancers usually say the deliverable is the source code. They rarely say when it changes hands, what counts as complete, or who holds the accounts while work is in progress. Three lines cover it:</p>
<ul>
<li>All source code and infrastructure are created in the client's accounts from day one. The vendor gets collaborator access, never ownership.</li>
<li>Intellectual property transfers on final payment. Until then the vendor grants a license to use the work in progress.</li>
<li>Final payment is due after the client has deployed a build themselves at least once.</li>
</ul>
<p>That third line is the useful test. If you have never deployed your own product, you have not really received it.</p>
<h2>A note on speed, from the other side of the table</h2>
<p>We build in the client's repo and deploy into the client's accounts. Day one is an empty repository with the client's name on it and a two-minute conversation about where the keys live. When the work ends, the client has a running product and every account it needs, and we can disappear without breaking anything. That is the point.</p>
<p>The numbers, since this post is also an introduction: a scoped first version takes 3–5 days. The same scope through an agency in Seoul usually runs about three months; we land near 1/20 of the price, not because the work is cheaper, but because there is no sales layer, no project manager, and no office between the person quoting and the person building.</p>
<p>Zero payment before a contract. A free code review of what you already have, so neither side is guessing. Deposit to start, balance after you have used the thing yourself. Contracts are direct with CPC AI Inc., a registered company in Seoul (business number 771-88-03539).</p>
<p>If you have half an app from Cursor, Lovable, Bolt or v0 and want to know what handover should look like, the honest answer takes 20 minutes of reading your repo. Send one line to <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a> — in your own language, we reply in it.</p>
<p>Studio (English): <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></p>
]]></content:encoded></item><item><title><![CDATA[The three features that were worth leaving out]]></title><description><![CDATA[A founder showed me his booking idea on a Tuesday. Places to stay, like every other listing site, except for one small group of people nobody was serving…]]></description><link>https://cpcai.hashnode.dev/the-three-features-that-were-worth-leaving-out</link><guid isPermaLink="true">https://cpcai.hashnode.dev/the-three-features-that-were-worth-leaving-out</guid><category><![CDATA[Startups]]></category><category><![CDATA[webdev]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sat, 14 Nov 2026 18:56:20 GMT</pubDate><content:encoded><![CDATA[<p>A founder showed me his booking idea on a Tuesday. Places to stay, like every other listing site, except for one small group of people nobody was serving well. He had a prototype, two weeks of evenings, and a list of three things he was sure the app needed before anyone could use it:</p>
<ol>
<li>chat between host and guest,</li>
<li>referral credits,</li>
<li>an admin dashboard so he could "manage things".</li>
</ol>
<p>He was not wrong that a finished product would eventually have all three. He was wrong about the order.</p>
<h2>The only question that decided the week</h2>
<p>I asked one thing: <em>who opens this first, and what do they have to be able to do in the first thirty seconds?</em></p>
<p>His answer: someone looking for a room, who needs to see it, pick a date, and put down a deposit.</p>
<p>None of the three features live in those thirty seconds. Each of them is a small product of its own:</p>
<ul>
<li><strong>Chat</strong> is not a text box. It is notifications, blocking, reporting, a record you can be asked for later, and someone to answer at 2am.</li>
<li><strong>Referral credits</strong> is money. Money that can be farmed, double-counted, and refunded.</li>
<li><strong>An admin dashboard</strong> is a second app with its own permissions and its own mistakes.</li>
</ul>
<p>Together: three weeks. Possibly five, and possibly never, because most projects die in the middle of their second week, not their first.</p>
<h2>What we shipped instead</h2>
<p>The ugly version. A stranger could open a link on a phone, see a room, pick a date, and pay a deposit. Hosts could add a room with photos and a price. Nothing else. It went live on Friday - not "ready for review", but actually reachable by a person who was not him.</p>
<p>The first bookings came in without chat, without credits, and with him answering questions by email like a human being.</p>
<p>Then the complaints arrived, and this is the part worth writing down: the complaints were about the photos, the price being unclear, and one confusing size description. Not one person asked for the features he had spent two weeks worrying about. The things real users wanted were things he could fix in an afternoon.</p>
<h2>The uncomfortable general rule</h2>
<p>Features are cheap to imagine and expensive to finish. An idea costs a sentence; a finished feature costs the boring half nobody pictures - the empty state, the error, the email, the edge case where someone does it twice.</p>
<p>Every app that never shipped is carrying a list of things nobody asked for, added by the person who was most afraid of launching.</p>
<p>So the "3 to 5 days" is not about working faster than an agency. It is about <strong>one decision made before anything is built</strong>: what "done this week" means, and what deliberately does not get built yet. That second list is the whole trick. Without it, a weekend of building turns into six months of almost.</p>
<h2>When this honestly works, and when it does not</h2>
<p>It works for the focused 80% of ideas people actually have: accounts, a list of things, a form, a payment, a notification. It works even better if you already started with an AI tool and stalled, because half the work is already on disk and only needs finishing.</p>
<p>It does not work for a bank integration, a two-sided marketplace with compliance, or a game whose content still needs months of design. Anyone who promises those in five days is guessing, and you will be the one paying for the guess.</p>
<p>I do this alone, from Seoul: a one-person studio, a direct contract, an invoice. No account manager between you and the person writing the code, so a question at 11pm gets an answer instead of a ticket number. The first conversation is free, including a look at whatever you already built.</p>
<ul>
<li>English studio page: <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></li>
<li>Email: <a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></li>
</ul>
<p>If you are holding a list of "must-haves" right now, try this: cross out everything that is not inside the first thirty seconds. What is left is a week. What you crossed out is next month's problem - and by then you will know whether anyone wants it at all.</p>
]]></content:encoded></item><item><title><![CDATA[Your app idea is stuck at 80%. In Seoul, one person just shipped three.]]></title><description><![CDATA[It is 2 a.m. in Seoul and someone is watching a tank shell fly. It arcs over a hill, catches the wind, and lands exactly on the other player's base…]]></description><link>https://cpcai.hashnode.dev/your-app-idea-is-stuck-at-80-in-seoul-one-person-just-shipped-three</link><guid isPermaLink="true">https://cpcai.hashnode.dev/your-app-idea-is-stuck-at-80-in-seoul-one-person-just-shipped-three</guid><category><![CDATA[showdev]]></category><category><![CDATA[AI]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sat, 14 Nov 2026 16:00:39 GMT</pubDate><content:encoded><![CDATA[<p>It is 2 a.m. in Seoul and someone is watching a tank shell fly.</p>
<p>It arcs over a hill, catches the wind, and lands exactly on the other player's base. Somewhere in another city, a real person just lost that round. The person who made the game wrote that physics, drew those tanks, and typed the store description himself. He is alone at his desk at 2 a.m. because that is when the servers are quiet and nobody is sending emails.</p>
<p>I am telling you about him because most people with an app idea get stuck in one very specific place — and it is not the place they think.</p>
<h2>The wall everyone hits at 80%</h2>
<p>You know the wall. You opened Cursor, or Lovable, or you typed a wish into ChatGPT, and for two days it was magic. Real screens appeared. A login page. A pretty green button that said Pay.</p>
<p>Then the magic stopped.</p>
<p>The app runs on your laptop but not on your phone. The payment button opens a page that says 401. The store reviewer sends back a screenshot with a red circle drawn on it. You close the laptop, promise yourself you will come back to it after the weekend, and the project goes into a folder called <code>final_v3_working</code>.</p>
<p>That folder is the most crowded graveyard in software. Not ideas that were bad — ideas that were 80% done.
The last 20% is not a feature. It is a hundred small frictions that only appear when a stranger touches your thing on a device you have never held: the keyboard that covers the button, the payment provider that refuses your country, the release build that crashes because a setting that existed on your laptop does not exist on a real phone.</p>
<h2>What one person in Seoul did with those same days</h2>
<p>CPC AI Inc. is a studio in Yeouido, Seoul. It is, honestly, one person.</p>
<p>In one stretch of days he:</p>
<ul>
<li>put a turn-based artillery game — <strong>Cannon King</strong>, with 48 collectible tanks, 53 story stages and live online battles — on <strong>Google Play</strong> and inside <strong>Toss</strong>, the payment app that half of Korea has sitting on the home screen of their phone;</li>
<li>runs <strong>CPC AI</strong>, a crypto app that explains <em>why</em> the market moved today instead of just showing you a red candle;</li>
<li>rebuilt his own studio website from a blank page to live, with the quote form sending emails to his inbox, in a single day.</li>
</ul>
<p>The game is not a demo. Real people downloaded it, real people lose to each other online at 2 a.m., and every one of those screens was drawn and coded by the same person who answers the emails.</p>
<h2>So how is that possible?</h2>
<p>There is no secret, and that is the interesting part.
<strong>He builds it himself.</strong> No agency, no account manager, no one "syncing with the team", no handover between three companies. The person who understands your idea is the person writing the code.</p>
<p><strong>The boring 70% is automated.</strong> Years of building his own pipeline means the repetitive parts — screens, forms, admin pages, deployment, store listings — get produced in hours by tools he controls. His hours go where a machine still loses: deciding what the thing should do, and checking that it actually does it on a real phone.</p>
<p><strong>He only takes work he can finish in days.</strong> A project that a typical agency quotes at 3–6 months and a five-figure invoice comes back at <strong>3–5 days</strong> and up to <strong>1/20 of the price</strong> — because nobody is paying for meeting rooms.</p>
<h2>The part that matters when it is your money</h2>
<p>Here is what I would want to know if I were handing my idea to a stranger on the internet.</p>
<p><strong>You do not pay to start.</strong> Quotes and consultations are free. At contract you pay a deposit, and the balance only after you have seen the thing working.</p>
<p><strong>You see real screens every day</strong>, and revisions are free until you are happy with it. Not a status report — the actual screens, on the actual device.
<strong>You are contracting with a registered Korean corporation</strong> (Business Reg. No. 771-88-03539) and you get an official invoice. Not an anonymous account in a chat server.</p>
<p><strong>If you are the one stuck at 80%</strong>, the first step is a free review of the code you already have. Nobody asks you to throw away the two weekends you already spent.</p>
<h2>What would you actually build?</h2>
<p>Not a unicorn. The small thing you have wanted for years.</p>
<p>The pilates studio you go to that still takes bookings through screenshots in a chat app. The shop with thirty products and a cart that works on a phone. The chatbot that answers "are you open on Sunday?" at midnight so you can sleep. The 20-second video ad for the thing you already sell, that you keep meaning to make.</p>
<p>Any one of those, said in a single line, is enough to start. You do not need a spec, a wireframe, or a technical vocabulary. You need one sentence and someone who will show you a screen this week.</p>
<h2>Where to look</h2>
<p>The studio's site is public, and the work is on it: <strong><a href="https://cpcai-studio.netlify.app/en/">cpcai-studio.netlify.app/en</a></strong>. The quote form is three fields — name, contact, one line about what you want — at <strong><a href="https://cpcai-studio.netlify.app/en/#inquiry">cpcai-studio.netlify.app/en/#inquiry</a></strong>. The same person who built the game reads it and replies within an hour. If you would rather write an email: <strong><a href="mailto:yyhak12345@gmail.com">yyhak12345@gmail.com</a></strong>.</p>
<p>Worst case, you get a free opinion on your idea from someone who actually ships. Best case, the folder called <code>final_v3_working</code> finally becomes an app on someone's phone — and you are the one watching a thing you made work at 2 a.m.</p>
]]></content:encoded></item><item><title><![CDATA[How I quote a rescue job in 30 minutes without reading the whole codebase]]></title><description><![CDATA[A triage process for half-finished AI-generated apps: what I ask for, what I run, and the three classes every project falls into.]]></description><link>https://cpcai.hashnode.dev/how-i-quote-a-rescue-job-in-30-minutes</link><guid isPermaLink="true">https://cpcai.hashnode.dev/how-i-quote-a-rescue-job-in-30-minutes</guid><category><![CDATA[AI]]></category><category><![CDATA[webdev]]></category><category><![CDATA[Freelancing]]></category><category><![CDATA[Productivity]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sat, 14 Nov 2026 07:01:47 GMT</pubDate><content:encoded><![CDATA[<p>People send me the same message about once a day now: "I built this with Lovable / Cursor / Bolt. It works locally. What would it cost to finish it?"</p>
<p>They expect me to read the repository. I don't, not at the quoting stage. Reading forty thousand generated lines is a week of unpaid work and it answers the wrong question. What decides whether a rescue is a three-day job or a three-week job is not the code style. It is how many unverified seams the project has.</p>
<p>Here is the process, in the order I run it. It takes about half an hour and it has been right often enough that I now quote from it.</p>
<h2>Step 1 — Ask for five things, not the repo</h2>
<p>Before I open anything:</p>
<ol>
<li><strong>The exact symptom, as a sentence.</strong> "Stripe works in test mode and returns 401 in live mode" is a quote I can make. "It doesn't work" is not.</li>
<li><strong>Where it runs now.</strong> localhost, a preview URL, a real domain, a store build. This single answer changes the size of the job more than the code does.</li>
<li><strong>What it is built with.</strong> Framework, database, auth provider, payment provider, host. Six words each.</li>
<li><strong>What "launched" means to them.</strong> Paying customers? A store listing? A demo for an investor? A rescue with no defined finish line has no end and no price.</li>
<li><strong>Whether they own the accounts.</strong> The database, the payment account, the domain, the store listing. If those belong to a contractor or a now-dead trial, that is the first work item, not the code.</li>
</ol>
<p>Missing answers here are already a finding. A builder who cannot say where the database lives does not have a deployable product yet, no matter how good the UI looks.</p>
<h2>Step 2 — Repo access, read the tree, not the files</h2>
<p>I clone it and look at structure: folder count, whether there is a <code>migrations</code> folder, whether there is a <code>prisma</code> or <code>drizzle</code> schema, whether there are tests, whether there is a single file over a thousand lines, whether secrets are in the tree, whether there is a README that matches reality.</p>
<p>Then three greps, which tell me most of what I need:</p>
<ul>
<li><code>process.env</code> — are variables referenced that do not exist in <code>.env.example</code>?</li>
<li><code>fetch(</code> and <code>axios</code> — how many external services are called, and is any of it called from the client?</li>
<li><code>id</code> and <code>user_id</code> — does authorization appear anywhere, or only in the UI?</li>
</ul>
<p>Generated code is usually well-organized and structurally dishonest. The tree lies less than the code.</p>
<h2>Step 3 — Run it, in three places</h2>
<ol>
<li><strong>Locally, following only the README.</strong> If I cannot get it running from the README in fifteen minutes, no future maintainer can either, and the count starts there.</li>
<li><strong>On a deployed preview with a clean environment.</strong> Missing env vars surface here.</li>
<li><strong>On a phone, on cellular data.</strong> Layout, cookies and third-party redirects all change.</li>
</ol>
<p>I do not fix anything during triage. Every fix I make during a quote is a fix I cannot price honestly later, and it hides the next failure.</p>
<h2>Step 4 — The four seams I test by hand</h2>
<p>Cheap tests, big signal:</p>
<ul>
<li><strong>Two accounts.</strong> Log in as A, try to read B's record by changing an id. If it returns data, the project needs an authorization pass on every endpoint. This is the most common serious finding and the one that decides whether I will take the job at all.</li>
<li><strong>A refresh and a cold start.</strong> Sessions that survive neither are not sessions.</li>
<li><strong>A live payment, then a refund.</strong> Not a test card. A webhook that was never verified in live mode is not a payment integration.</li>
<li><strong>The build, from a clean clone on a machine that is not theirs.</strong> Half of "works on my machine" is a missing lockfile.</li>
</ul>
<h2>Step 5 — Classify, then price</h2>
<p>Every project lands in one of three classes.</p>
<p><strong>Class A — days, not weeks.</strong> The architecture is sound; what is missing is integration and polish: env validation, deploy and domain, webhook verification, a handful of bugs, mobile layout. This is a three-to-five day finish and I quote it as one price with a fixed finish line.</p>
<p><strong>Class B — one rebuilt layer.</strong> The app is real but one layer cannot be rescued: auth written from scratch instead of through a provider, or data access with no ownership model at all, or a payment flow that needs to be re-architected because it grants access from the redirect. I rebuild that layer, keep everything else, and quote it as two phases so the client sees what the rebuild costs before committing to it.</p>
<p><strong>Class C — the honest no.</strong> No repository history, no ownership of the accounts, or a product whose requirements are still moving weekly. In those cases the cheapest route is usually not me; it is to keep building for another month and come back when the shape stops changing. I say so, and I say why. A quote I cannot honour costs the client more than a lost job costs me.</p>
<h2>What the quote itself contains</h2>
<p>Not an hourly estimate. A finish line:</p>
<ul>
<li>what "done" means, in a sentence the client wrote</li>
<li>the list of seams to be verified, against the four above</li>
<li>what I need from them, and by when</li>
<li>the price, the deposit at contract, and the balance after they check the result</li>
<li>the revision scope — what counts as a revision and what counts as new work</li>
</ul>
<p>The last two lines are why the process works. Rescue work goes wrong when "finish it" has no edge. Written down, it has one.</p>
<h2>The uncomfortable part</h2>
<p>Most of the "last 20%" is not code. It is decisions nobody made: which host is canonical, who owns the payment account, what happens when a webhook arrives twice, who can read which row. A generator cannot make those decisions for you, and neither can a quote that only counts files.</p>
<p>If you are sitting on an app that has been two weeks from launch for two months, the useful next hour is not more prompting. It is answering the five questions in step 1 on paper. If you would rather someone did it with you, that is the work I do:</p>
<p><strong><a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></strong> — free review first, then a quote with a finish line. One-person studio, direct contract, no agency in between.</p>
]]></content:encoded></item><item><title><![CDATA[Running a one-person studio with no on-call rotation]]></title><description><![CDATA[No one to page: the three alerts a one-person studio keeps, what to delete, and why idempotency is an ops tool, not a nice-to-have.]]></description><link>https://cpcai.hashnode.dev/running-a-one-person-studio-with-no-on-call-rotation</link><guid isPermaLink="true">https://cpcai.hashnode.dev/running-a-one-person-studio-with-no-on-call-rotation</guid><category><![CDATA[Devops]]></category><category><![CDATA[automation]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sat, 14 Nov 2026 04:14:54 GMT</pubDate><content:encoded><![CDATA[<p>There is no one to page. That is the whole design constraint of a one-person software business: if something breaks at 3am, the alert either wakes me or nobody finds out until a user leaves.</p>
<p>I run a live web app, a mobile game on two stores, and client projects. Here is the monitoring I actually keep, and what I deleted because it only made noise.</p>
<h2>The three things worth waking up for</h2>
<p>Everything else can wait until morning. I keep exactly three wake-me conditions:</p>
<ol>
<li><strong>The core request path is down.</strong> Not slow — down. A health endpoint that touches the database and one critical dependency, checked from outside.</li>
<li><strong>Writes are failing.</strong> Reads failing is bad; writes failing silently is worse, because users lose data and do not tell you.</li>
<li><strong>Money-touching code raised an exception.</strong> Purchase verification, coupon redemption, subscription changes.</li>
</ol>
<p>Anything else — CPU, memory, a 500 rate under 1% — goes into a daily digest.</p>
<h2>What I deleted</h2>
<ul>
<li><strong>Uptime checks on every route.</strong> Six checks that all fail together when the server is down produce six pages for one incident.</li>
<li><strong>Log-based alerts on error count.</strong> One broken client with a retry loop can produce 40k errors and mean nothing.</li>
<li><strong>Disk and memory thresholds.</strong> They fired during deploys, which is normal.</li>
</ul>
<p>Deleting alerts is a feature. An alert I learned to ignore is worse than no alert, because it teaches me to skim the one that matters.</p>
<h2>Idempotency as an operational tool</h2>
<p>The most useful ops decision was in code, not monitoring: every write path that can be retried carries an idempotency key.</p>
<ul>
<li>Payment webhooks: the same event delivered twice changes nothing.</li>
<li>Coupon redemption: per account, enforced at the database, not in the UI.</li>
<li>Video render jobs: keyed by input hash, so a retry returns the existing artifact.</li>
</ul>
<p>With those in place, most incidents become retry it later instead of did we double-charge someone.</p>
<h2>Backups you have actually restored</h2>
<p>I test a restore monthly from the nightly backup into a scratch database. Untested backups are a belief, not a backup. The test takes ten minutes and has caught a missing table once.</p>
<h2>The ops work I handed off</h2>
<p>I am the only developer, so the non-engineering half — release notes, store listings, support triage, publishing to five channels in three languages — is handled by an AI assistant working from written briefs. It drafts, files, and posts where I told it to; I keep the decisions and anything that spends money or speaks in my name.</p>
<p>If you are solo and drowning in the operational tail: write the brief once, then let something else carry it. (I use Pickle for this; invite code SJWDRVNKFDSY — <a href="https://pickle.com/i/SJWDRVNKFDSY">https://pickle.com/i/SJWDRVNKFDSY</a>)</p>
<hr />
<p>I build and ship apps, games, and client MVPs from Seoul: <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></p>
]]></content:encoded></item><item><title><![CDATA[From headline to a 27-second vertical video: the unattended render pipeline behind a crypto market app]]></title><description><![CDATA[How I built an unattended pipeline that turns a verified crypto headline into a 27-second vertical video: Playwright frames, ffmpeg, cached TTS]]></description><link>https://cpcai.hashnode.dev/from-headline-to-a-27-second-vertical-video-the-unattended-render-pipeline-behind-a-crypto-market-app</link><guid isPermaLink="true">https://cpcai.hashnode.dev/from-headline-to-a-27-second-vertical-video-the-unattended-render-pipeline-behind-a-crypto-market-app</guid><category><![CDATA[automation]]></category><category><![CDATA[AI]]></category><dc:creator><![CDATA[Youngjun Yoon]]></dc:creator><pubDate>Sat, 14 Nov 2026 02:09:24 GMT</pubDate><content:encoded><![CDATA[<p>Crypto news moves in minutes, and a 27-second vertical video is a format that can ride it — but only if the loop from "headline confirmed" to "file on disk" runs without me. This is the pipeline I built and still run for a one-person market-analysis app, including the parts that turned out to matter.</p>
<h2>The shape of the loop</h2>
<p>TRIGGER → FACT-CHECK → SCRIPT → VOICE → RENDER → PUBLISH</p>
<p>Each step is a separate process and every boundary is a file on disk. No queue, no orchestration service, no shared in-memory state. If step 4 fails at 3 a.m., I rerun step 4 alone with the artifacts from steps 1–3.</p>
<h2>1. Trigger and fact-check</h2>
<p>A watcher polls exchange and on-chain sources plus a couple of news feeds. The rule for turning an item into a video is deliberately narrow: a number that moved materially (liquidation size, funding flip, a large exchange deposit) <strong>and</strong> a source I can name.</p>
<p>The most important code in this step is a rejection: if the number cannot be checked against a second source, no video is produced. A wrong number with a clean render is worse than no post at all.</p>
<h2>2. Script: a fixed skeleton with slots</h2>
<p>The script is a template, not free-form generation:</p>
<pre><code>hook:  {asset} {direction} {magnitude} in {window}
body:  what happened / why it matters / what to watch next
cta:   one fixed line, 1.3s of read time
</code></pre>
<p>Because the structure never changes, the pipeline stays testable: every render must contain a hook, exactly one body claim and the same CTA. If the template stops being boring, the automation stops being reliable.</p>
<h2>3. Voice</h2>
<p>TTS per sentence, cached by text hash. A re-render reuses existing audio, which takes a re-run from minutes to seconds and makes fixing a typo cheap.</p>
<h2>4. Render</h2>
<p>An HTML template at 1080×1920, 9:16, 30fps, driven frame by frame by a headless browser. Frames are written to a temp directory; ffmpeg muxes them with the voice track, a music bed and a fixed sub-line. Output is H.264 CRF 17, about 27 seconds long.</p>
<p>Two things I learned the hard way:</p>
<ul>
<li><strong>Frames to disk, then encode.</strong> Streaming frames straight into ffmpeg looks elegant and dies on the first long GC pause.</li>
<li><strong>No network calls inside the template.</strong> A font fetched at render time is a video that fails silently at 3 a.m.</li>
</ul>
<h2>5. Publish, with proof attached</h2>
<p>Every output lands with a manifest: asset, time window, both sources, render hash. The text version goes to the blog and the video is attached. The manifest is what lets me answer "where did that number come from" three weeks later — and it is the difference between a pipeline and a content farm.</p>
<h2>What it costs</h2>
<p>One render is about a minute of CPU and roughly $0.05 of TTS. Compute was never the bottleneck. Deciding what is worth saying is.</p>
<h2>What I still do by hand</h2>
<p>The decision to post. Everything up to the file is mechanical; whether today's move deserves a video is judgment, and it is the one place a wrong call actually costs credibility.</p>
<p>If you are building something similar — or you have a half-finished pipeline that only runs when you sit next to it — I write about this and build it for clients from a one-person studio in Seoul: <a href="https://cpcai-studio.netlify.app/en/">https://cpcai-studio.netlify.app/en/</a></p>
]]></content:encoded></item></channel></rss>