Spec-ops for non-coders. $400/mo outperforms $1,000/mo.
You don't need to code. You need to spec.
A structured development methodology that lets non-technical founders build working systems using AI. Tested on real projects (see case studies), not a theoretical framework.
- Claude Opus = Your CTO (writes specs, cuts tasks, reviews code)
- Codex CLI = Your engineer (writes code, runs tests, reports back)
- You = The product owner (says what you want in plain language)
No coding required. Seriously.
Claude MAX plan charges Opus tokens at ~10x the rate of lighter models. When you use Opus for everything — including writing for-loops and fixing CSS — you're paying PhD rates for manual labor.
5x-cto splits the work:
| Role | Who | Token Budget |
|---|---|---|
| Thinking (specs, review, acceptance) | Opus (~15-20%) | Your Claude MAX |
| Building (code, debug, tests) | Codex (~80-85%) | Separate Codex budget |
Result: Your Claude MAX subscription stretches significantly further because 80% of the work runs on a separate budget.
How we estimate "5x": On Claude MAX $200/mo, Opus gets ~24-40 hours/week vs Sonnet's ~240-480 hours/week (~10x difference in allocation burn rate). With this methodology, Opus handles only ~20% of work (specs + review). 100% / 20% = 5x. This is a theoretical upper bound based on Anthropic's published rate limits, not a controlled experiment. Your actual ratio depends on task complexity and rework cycles. Conservative range: 4-7x.
| Setup | Monthly Cost | Capacity | Quality |
|---|---|---|---|
| 5x Claude MAX (all Opus) | $1,000 | Not enough | 5/5 |
| 5x Claude MAX (all Sonnet) | $1,000 | 10x more | 4/5 |
| 1x MAX + 1x Codex (5x-cto) | $400 | >= all-Opus | 5/5 |
Same quality at the Opus-thinking layer. 60% less spend on Claude MAX.
This methodology is your technical ability.
It asks the right questions, breaks work into the right steps, and checks quality at every gate. You focus on what problem you're solving — the process handles the engineering discipline.
Projects built with this methodology (by a non-technical founder):
- API system integrating 30+ external services
- Multi-language AI chatbot
- Certification & examination platform
- E-commerce with Stripe payments, PDF certificates, QR verification
These systems are deployed and running. See case studies for scored evidence.
Honest disclaimer: This methodology handles structure, discipline, and quality checks. It does not replace domain expertise, security audits, or production operations knowledge. The scorer is heuristic-based (see limitations in
scripts/score.sh).
You show up 3 times. AI handles the rest.
You AI
| |
+- "I want X" -----------> Phase R: Structures your requirements
| |
+- "Yes, that's right" --> Phase S: Writes technical spec
| Phase 1: Cuts work into cards
| Phase 2-5: Auto-runs ALL cards
| +- Card 1 -> Build -> Review -> PASS
| +- Card 2 -> Build -> Review -> PASS
| +- Card 3 -> Build -> Review -> PASS
| +- ...
| Phase 5.5: Total System Review
| |
+- Reviews final report <--+
Every card goes through:
- Scope control — Only touches files in scope. No side effects.
- Acceptance criteria — Pass/fail, not "looks good"
- Independent verification — Opus re-runs checks, doesn't trust self-reports
- Mandatory 3 checks — Tests pass, error handling covered, commit has Card ID
- Rework loop — Fail? Fix and resubmit. 3 failures = blocked, escalate.
After ALL cards complete, Total Acceptance:
- End-to-end flow verification
- Cross-module integration check
- Architecture consistency review
- Security audit
- Deploy readiness assessment
- Automated scoring (see
scripts/score.sh)
- Claude Code with MAX plan ($200/mo) — set to Opus model
- OpenAI Codex CLI with Pro plan ($200/mo)
# Copy the skill to your Claude Code skills directory
cp -r skills/5x-cto ~/.claude/skills/
# Verify
ls ~/.claude/skills/5x-cto/SKILL.mdOpen Claude Code and say:
"I want to build a bookmark API with auth, CRUD, and search"
The skill activates automatically. You'll be guided through requirements, then it runs.
After your pipeline completes:
bash scripts/score.sh /path/to/your/projectReal projects, real scores, built by a non-technical founder:
| Date | Project | Cards | Score | Highlights |
|---|---|---|---|---|
| 2026-03-16 | AI Academy | 9/9 | 92/100 | Stripe + PDF certs + i18n + 10 tests |
| 2026-03-16 | Boss Dashboard | 4/4 | 85/100 | Real-time status + 3-lang UI |
Q: Do I really not need to code? A: Correct. You need to know what problem you're solving and who it's for. The methodology handles the rest.
Q: What if a card fails 3 times? A: It's marked "blocked." The system skips it, continues with other cards, and reports the blocker. You decide: re-scope, simplify, or have Opus take over.
Q: Is 5x really accurate? A: Theoretical estimate based on Anthropic's published rate limits. Opus handles ~20% of work (specs + review), rest goes to Codex. 100%/20% = 5x. Not independently measured — your results depend on task complexity and rework rate. See methodology section for details.
Q: Can I use this with other AI tools? A: The methodology is model-agnostic. The skill is written for Claude + Codex, but the principles (requirement -> spec -> card -> build -> review) work with any AI coding tool.
你不需要會寫程式。你只需要會說你要什麼。
一套有結構的開發方法論,讓不會寫程式的人也能用 AI 做出可運行的系統。已在真實專案中測試過(見案例)。
- Claude Opus = 你的技術長(寫規格、切任務、驗收程式碼)
- Codex CLI = 你的工程師(寫程式、跑測試、回報進度)
- 你 = 產品負責人(用人話說你要什麼就好)
不用寫程式。真的。
Claude MAX 方案中,Opus 消耗額度的速度是一般模型的 ~10 倍。當你用 Opus 做所有事情 — 包括寫迴圈和改 CSS — 等於請博士生搬磚。
5x-cto 把工作分開:
| 角色 | 誰做 | Token 預算 |
|---|---|---|
| 思考(規格、驗收、審查) | Opus(~15-20%) | 你的 Claude MAX |
| 建造(寫碼、除錯、測試) | Codex(~80-85%) | Codex 獨立額度 |
結果: 你的 Claude MAX 訂閱效率提升 5 倍,因為 80% 的工作走的是另一個預算。
| 方案 | 月費 | 容量 | 品質 |
|---|---|---|---|
| 5 個 Claude MAX(全用 Opus) | $1,000 | 不夠用 | 5/5 |
| 5 個 Claude MAX(全用 Sonnet) | $1,000 | 多 10 倍 | 4/5 |
| 1 個 MAX + 1 個 Codex(5x-cto) | $400 | >= 全 Opus | 5/5 |
Opus 思考層的品質不變。Claude MAX 花費省 60%。
這個方法論就是你的技術能力。
它幫你問對的問題、拆對的步驟、卡對的品質。你專注於你要解什麼問題 — 流程處理工程紀律。
用這套方法論建的專案(由一位沒有技術背景的創業者完成):
- 整合 30+ 外部服務的 API 系統
- 多語言 AI 聊天機器人
- 認證與考試平台
- 含 Stripe 付費、PDF 證書、QR 驗證的電商系統
這些系統已部署並運行中。見案例了解有分數的證據。
誠實聲明: 這個方法論處理結構、紀律和品質檢查。它不能取代領域知識、安全稽核或生產維運經驗。評分器是啟發式的(見
scripts/score.sh中的限制說明)。
你只需要出現 3 次,AI 做剩下的。
你 AI
| |
+- 「我要 X」 ------------> Phase R: 結構化你的需求
| |
+- 「對,就是這個」 ------> Phase S: 寫技術規格
| Phase 1: 切成工作卡片
| Phase 2-5: 自動跑完所有卡片
| +- 卡 1 → 施工 → 驗收 → 通過
| +- 卡 2 → 施工 → 驗收 → 通過
| +- 卡 3 → 施工 → 驗收 → 通過
| +- ...
| Phase 5.5: 總驗收
| |
+- 看總驗收報告 <-----------+
每張卡都必須通過:
- 範圍控制 — 只碰該碰的檔案,零副作用
- 驗收標準 — 通過/不通過,不是「看起來可以」
- 獨立驗證 — Opus 自己重跑檢查,不相信自述報告
- 強制三項檢查 — 測試通過、錯誤處理覆蓋、Commit 帶卡號
- 返工循環 — 不合格?修完再交。連續 3 次不過 = 擱置,上報
所有卡完成後,進行總驗收:
- 端對端流程驗證
- 跨模組整合檢查
- 架構一致性審查
- 安全審計
- 部署就緒評估
- 自動跑分(見
scripts/score.sh)
真實專案、真實分數,由一個沒有技術背景的創業者完成:
| 日期 | 專案 | 卡片 | 分數 | 亮點 |
|---|---|---|---|---|
| 2026-03-16 | AI 學院 | 9/9 | 92/100 | Stripe + PDF 證書 + 三語 + 10 測試 |
| 2026-03-16 | 老闆儀表板 | 4/4 | 85/100 | 即時狀態 + 三語 UI |
コードを書く必要はありません。必要なのは「何を作りたいか」を伝えることだけ。
技術的な知識がない人でも、AIを使って動作するシステムを構築できる、構造化された開発方法論です。実プロジェクトで検証済み(ケーススタディ参照)。
- Claude Opus = あなたのCTO(仕様書を書き、タスクを分割し、コードをレビュー)
- Codex CLI = あなたのエンジニア(コードを書き、テストを実行し、結果を報告)
- あなた = プロダクトオーナー(やりたいことを普通の言葉で伝えるだけ)
コーディング不要。本当です。
Claude MAXプランでは、Opusのトークン消費速度は軽量モデルの約10倍。Opusでforループの記述やCSS修正まで全てやるのは、博士号を持つ人に引越し作業をさせるようなものです。
5x-ctoは作業を分割します:
| 役割 | 担当 | トークン予算 |
|---|---|---|
| 思考(仕様・レビュー・検収) | Opus(約15-20%) | あなたのClaude MAX |
| 構築(コード・デバッグ・テスト) | Codex(約80-85%) | Codex独自の予算 |
結果: Claude MAXサブスクリプションの効率が5倍に。80%の作業は別予算で実行されるため。
| 構成 | 月額 | 容量 | 品質 |
|---|---|---|---|
| Claude MAX x5(全てOpus) | $1,000 | 足りない | 5/5 |
| Claude MAX x5(全てSonnet) | $1,000 | 10倍多い | 4/5 |
| MAX x1 + Codex x1(5x-cto) | $400 | 全Opus以上 | 5/5 |
Opus思考レイヤーの品質はそのまま。Claude MAXの支出は60%削減。
この方法論があなたの技術力になります。
正しい質問をし、正しいステップに分解し、各ゲートで品質をチェック。あなたは解決したい問題に集中するだけ — プロセスがエンジニアリング規律を担保します。
この方法論で構築したプロジェクト(技術未経験の創業者による):
- 30以上の外部サービスと連携するAPI
- 多言語AIチャットボット
- 資格認定・試験プラットフォーム
- Stripe決済・PDF証明書・QR認証付きEC
これらのシステムは実際にデプロイ・稼働中です。ケーススタディでスコア付きの証拠をご確認ください。
正直な免責事項: この方法論は構造・規律・品質チェックを扱います。ドメイン知識、セキュリティ監査、本番運用の経験を代替するものではありません。スコアラーはヒューリスティックベースです(
scripts/score.sh内の制限事項を参照)。
あなたが登場するのは3回だけ。残りはAIが処理します。
あなた AI
| |
+- 「Xが欲しい」 ---------> Phase R: 要件を構造化
| |
+- 「そう、それ」 --------> Phase S: 技術仕様を作成
| Phase 1: 作業カードに分割
| Phase 2-5: 全カードを自動実行
| +- カード1 → 実装 → 検収 → 合格
| +- カード2 → 実装 → 検収 → 合格
| +- カード3 → 実装 → 検収 → 合格
| +- ...
| Phase 5.5: 総合検収
| |
+- 最終レポートを確認 <------+
各カードは以下を通過する必要があります:
- スコープ制御 — 対象ファイルのみ変更。副作用ゼロ
- 受入基準 — 合格/不合格。「良さそう」は不可
- 独立検証 — Opusが自らチェックを再実行。自己申告を信用しない
- 必須3項目チェック — テスト合格・エラーハンドリング・Commitにカード番号
- やり直しループ — 不合格?修正して再提出。3回失敗 = ブロック、エスカレーション
全カード完了後、総合検収を実施:
- エンドツーエンド検証
- モジュール間統合チェック
- アーキテクチャ一貫性レビュー
- セキュリティ監査
- デプロイ準備評価
- 自動スコアリング(
scripts/score.sh参照)
実プロジェクト、実スコア、技術未経験の創業者が構築:
| 日付 | プロジェクト | カード | スコア | ハイライト |
|---|---|---|---|---|
| 2026-03-16 | AIアカデミー | 9/9 | 92/100 | Stripe + PDF証明書 + 三言語 + テスト10件 |
| 2026-03-16 | ボスダッシュボード | 4/4 | 85/100 | リアルタイム状態 + 三言語UI |
- Claude Code + MAX plan ($200/mo)
- OpenAI Codex CLI + Pro plan ($200/mo)
cp -r skills/5x-cto ~/.claude/skills/
ls ~/.claude/skills/5x-cto/SKILL.mdOpen Claude Code and say what you want to build:
"I want to build a bookmark API with auth, CRUD, and search"
「我要做一個書籤 API,有登入、新增刪除、搜尋功能」
「ブックマークAPIを作りたい。認証、CRUD、検索機能付き」
bash scripts/score.sh /path/to/your/project- YES.md (
sstklen/yes.md) — AI governance: safety gates, evidence rules, anti-slack detection. 5x-cto handles the pipeline, YES.md keeps AI honest throughout. Recommended for production use. - AI.MD (
sstklen/ai-md) — Convert your CLAUDE.md to AI-native format so rules actually get followed.
MIT
Built by a non-technical founder in rural Japan. 由一位住在日本鄉下的非技術創業者打造。 日本の田舎に住む非技術者の創業者が構築。