Auto Project Builder
hongmacho/auto-project-builder
Constrói projetos Next.js+shadcn/ui+SQLite automaticamente, do zero ao deploy
Ver no GitHubSkill do Claude Code que constrói autonomamente projetos web completos com Next.js, shadcn/ui e SQLite a partir de ideias geradas. Executa pipeline integrado de pesquisa, scoring de ideias, análise competitiva, build, QA automático, geração de README e push para GitHub — tudo sem intervenção humana.
Instalação
npx convez-ai hongmacho/auto-project-builder1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar e construir aplicações web completas com validação automática
- Validar ideias de startup contra YC scorecard e mercado competitivo
- Acelerar prototipagem rápida com stack moderno pré-configurado
- Automatizar pesquisa de tendências e deduplicação de projetos similares
Veredito

Solução robusta e opinada para geração de projetos web full-stack com rigor de QA integrado. Excelente para prototipagem rápida e validação de conceitos, mas requer entendimento da stack (Next.js/TypeScript/SQLite) para customizações profundas.
README
A tradução para o português deste README ainda está na fila.
auto-project-builder
Finish working projects with zero human intervention. A project with build errors, type errors, or lint errors is not "done."
A Claude Code skill that runs a fully autonomous pipeline: interactive setup → trend research → idea scoring → competitive analysis → GitHub dedup → build → auto QA loop → README generation → GitHub push → report → enhancement guide.
How It Works
/auto-project-builder
│
▼
┌─────────────────────────────────────────────────────────┐
│ Phase -0.5 OMC / ECC Environment Detection (silent) │
│ → OMC_MODE = "omc" / "ecc" / "none" │
└──────────────────────────┬──────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ Phase -1 Checkpoint check + Interactive setup (up to 5)│
│ Q1 Platform → Q1.5 Have an idea? → Q2 Tech Stack │
│ → Q3 Service Type* → Q4 Count* (* skipped if Q1.5=yes)│
└──────────────────────────┬──────────────────────────────┘
│
┌──────────────────┴──────────────────┐
▼ ▼
┌───────────────┐ ┌──────────────────┐
│ Phase 0-A │ │ Phase 0-B │
│ Context7 │ (parallel) │ Trend Research │
│ Stack Docs │ │ Product Hunt / │
│ │ │ GitHub Trending │
└───────┬───────┘ └────────┬─────────┘
└──────────────┬────────────────────┘
▼
┌──────────────────────────────────────────────────┐
│ Phase 1 Idea Generation (idea-generator skill) │
│ · Pain mining: Reddit/HN/AppStore/GitHub Issues │
│ · 3-role adversarial eval: planner+architect+ │
│ critic (YC-style fatal-flaw-first) │
│ · User-idea path: use provided idea directly │
└──────────────┬───────────────────────────────────┘
▼
┌──────────────────────────────┐
│ Phase 1.3 Score Review │
│ 15-pt YC scorecard │
│ GO / CONDITIONAL / NO-GO │
│ → auto-replace NO-GO ideas │
│ (skipped for user's idea) │
└──────────────┬───────────────┘
▼
┌──────────────────────────────┐
│ Phase 1.5 GitHub Dedup │
│ Skip ideas already built │
└──────────────┬───────────────┘
▼
┌──────────────────────────────────────────────────┐
│ Phase 2 Build Loop (repeated per project) │
│ │
│ 2-1 Write PRD │
│ 2-2 Write ROADMAP + Sprint plan │
│ 2-3 Implement │
│ 2-4 ★ Auto QA loop (tsc / lint / build) │
│ └─ on fail: build-error-resolver ×3 │
│ 2-5 Auto-generate README │
│ 2-6 GitHub push │
│ 2-7 macOS completion notification │
│ 2-8 Save checkpoint │
│ 2-9 Generate enhancement guide (5 categories) │
│ 2-10 Update user-suggest.html │
└──────────────┬───────────────────────────────────┘
▼
┌──────────────────────────────────────┐
│ Phase 3 Reports │
│ {YYYYMMDDHHmm}_report.html │
│ overview.html (append-only) │
└──────────────────────────────────────┘
Installation
curl -fsSL https://raw.githubusercontent.com/hongmacho/auto-project-builder/main/install.sh | bash
This installs both required skills in one command:
~/.claude/skills/auto-project-builder/SKILL.md~/.claude/skills/idea-generator/SKILL.md
Restart Claude Code — the skills will appear in /skills.
Usage
/auto-project-builder
The skill asks up to 5 questions at startup, then runs fully autonomously.
| # | Question | Options |
|---|---|---|
| 1 | Platform | Web / Mobile App / CLI / Auto / Custom |
| 1.5 | Have an idea? | Auto (Claude generates from trends) / Direct input |
| 2 | Preferred Tech Stack | Platform-specific options + None (per-idea auto-pick) / Custom |
| 3 | Service Type * | Productivity / Content / Commerce / Education / … (multi-select) + Auto |
| 4 | Count * | 1–10 / Auto (= floor(categories × 2–3), max 10) |
* Questions 3 and 4 are skipped automatically when you provide your own idea at Q1.5 — PROJECT_COUNT is set to 1.
Selecting Auto / None on any question lets Claude decide autonomously based on trend data. The chosen value and reasoning are always announced before proceeding.
Features
Checkpoint Resume
Progress is saved to .auto-project-builder-checkpoint.json after each completed project. If the run is interrupted, the next invocation offers to resume from where it left off — completed projects are skipped automatically.
High-Quality Idea Generation (idea-generator skill)
Phase 1 delegates to a dedicated idea-generator skill that starts from real user pain, not trending technology. The process:
-
Multi-source pain mining (4 parallel lanes)
- Reddit: "I wish there was" / "why is there no" complaints
- Hacker News: "Ask HN: Is there a tool for..." threads
- App Store 1-star reviews: repeated frustration patterns in competing apps
- GitHub Issues: unresolved feature requests in popular repos
-
Pain-to-idea transformation — every idea is backed by real user quotes, not speculation
-
3-role adversarial evaluation (run in parallel)
- Planner: value proposition, killer feature, target persona
- Architect: technical feasibility, risks, MVP sprint estimate
- Critic (YC-style): fatal flaws first, market size, defensibility, monetization
-
15-point YC scorecard + go/no-go verdict
| Criterion | 1 pt | 2 pts | 3 pts |
|---|---|---|---|
| Pain strength | Speculation only | Indirect evidence | 5+ direct user quotes |
| Market size | Niche few | Thousands | Hundreds of thousands+ |
| Originality | Obvious clone | Improved clone | Novel combination |
| Feasibility | Stack mismatch | Possible but risky | Perfect stack fit |
| Fatal flaw penalty | — | — | −1 to −3 pts |
Verdicts: GO (≥11 pts) · CONDITIONAL (8–10 pts) · NO-GO (≤7 pts, auto-replaced)
GitHub Dedup
Runs gh repo list and filters out ideas that are too similar to existing repositories, then auto-generates replacements.
Auto QA Loop (Core)
After implementation, stack-specific validation commands are executed.
| Stack | QA Commands |
|---|---|
| TypeScript (Web / CLI) | tsc --noEmit && lint && build |
| Expo (App) | expo-doctor && tsc --noEmit |
| Python | mypy && pytest |
| Go | vet && build && test |
| Rust | check && clippy && test |
On failure, errors are handed to the build-error-resolver agent for automatic fixing. Up to 3 attempts. If all 3 fail, Nice-to-have features are stripped and only Must-have features are re-implemented. If that still fails, the project is marked SKIP and the next project begins.
Auto README Generation
After each QA pass, a README.md is generated for the project containing: feature list, tech stack table, prerequisites, installation steps, and usage commands.
Completion Notifications
A macOS notification is fired after each project completes so you can step away during long runs.
Per-Idea Tech Stack
When None is selected for the tech stack question, no stack is locked at setup time. Instead, each idea receives its own optimal stack at Phase 1 planning time based on the idea's requirements and current trends. This allows a single run to produce projects in Next.js, SvelteKit, and Remix simultaneously without being constrained to one stack up front.
Enhancement Guide (user-suggest.html)
After each project completes, the skill generates a five-category enhancement guide saved as projects/{slug}/user-suggest.html — one standalone file per project. Five projects → five independent files.
| Category | Contents |
|---|---|
| Quick Wins | Small improvements deployable within a day |
| Feature Enhancements | Mid-size features to add in the next sprint |
| Tech Improvements | Performance, security, and architecture upgrades |
| Growth Strategies | User acquisition, SEO, viral loops, partnerships |
| Monetization Ideas | Revenue models suited to the project type |
Raw suggestion data is also saved to report_data/{slug}_suggestions.json for programmatic use.
OMC Integration (oh-my-claudecode)
When running inside oh-my-claudecode, the skill automatically detects the environment at startup (Phase -0.5) and upgrades each phase with specialized agent orchestration.
Auto-Detection
| Detected | OMC_MODE | Agent Strategy |
|---|---|---|
oh-my-claudecode:* skills | "omc" | Full OMC orchestration |
everything-claude-code:* skills | "ecc" | ECC agent suite |
| Neither | "none" | Built-in agents |
Phase 1: Pain-Driven Idea Generation with idea-generator
When OMC_MODE = "omc", the idea-generator skill runs three agent roles in parallel for maximum idea quality:
| Role | Responsibility |
|---|---|
planner | Value proposition, killer feature, target persona, 6-month survival scenario |
architect | Technical feasibility score, implementation risks, MVP sprint count |
critic | YC-style fatal-flaw hunting, competitor analysis, monetization path, go/no-go |
This replaces simple trend-based generation with evidence-backed, adversarially validated ideas.
Phase 2: Full Autonomous Orchestration with autopilot
Optionally hand off all of Phase 2 (PRD → ROADMAP → implement → QA → README → push) to autopilot for zero-touch execution:
Skill("oh-my-claudecode:autopilot",
prompt="Complete {N} projects end-to-end.
PLATFORM / TECH_STACK / PRD path: projects/{slug}/docs/PRD.md
Done when: zero build errors + 80%+ coverage + README + GitHub push")
Post-Build Quality Loop with ralph
After implementation, ralph applies a PRD-driven acceptance-criteria loop until quality is guaranteed.
Auto-triggered when:
- QA failed 2+ times
- Idea scored ≤ 7/15 (CONDITIONAL verdict)
- Code review found 3+ HIGH issues
Criteria ralph enforces:
- Test coverage ≥ 80%
- Zero
tsc / lint / builderrors - Explicit error handling at every boundary
- Functions / components ≤ 200 lines
- All PRD Must-have features verified working
Can also be applied optionally to any completed project for uniform quality uplift.
Supported Tech Stacks
Web
| Option | Stack |
|---|---|
| ① Recommended | Next.js 14+ App Router · shadcn/ui · Drizzle ORM · SQLite |
| ② | Nuxt 3 · Tailwind CSS · PGlite |
| ③ | SvelteKit · shadcn-svelte · Drizzle ORM · SQLite |
| ④ | Remix · shadcn/ui · Drizzle ORM · SQLite |
| ⑤ | Custom input |
Mobile App
| Option | Stack |
|---|---|
| ① Recommended | React Native + Expo · expo-sqlite |
| ② | Flutter 3 · Dart · sqflite |
| ③ | Auto (Claude picks per idea based on trend data) |
| ④ | Custom input |
CLI
| Option | Stack |
|---|---|
| ① Recommended | Node.js · TypeScript · Commander.js · SQLite |
| ② | Python 3.12 · Typer / Click · SQLite |
| ③ | Go 1.22 · Cobra · SQLite |
| ④ | Rust · Clap · ratatui · SQLite |
| ⑤ | Custom input |
Output
{working directory}/
├── projects/
│ └── {slug}/
│ ├── docs/
│ │ ├── PRD.md
│ │ └── ROADMAP.md
│ ├── README.md ← auto-generated
│ ├── user-suggest.html ← per-project enhancement guide (standalone)
│ └── ... (source code)
├── report_data/
│ ├── {slug}_log.json
│ └── {slug}_suggestions.json ← enhancement guide data per project
├── .auto-project-builder-checkpoint.json ← exists during run, deleted on completion
├── {YYYYMMDDHHmm}_report.html ← per-run report, timestamped to minute (e.g. 202605242157_report.html)
└── overview.html ← cumulative catalog, append-only across all runs
{YYYYMMDDHHmm}_report.html
A new timestamped report is created for every run (hour + minute included, e.g. 202605242157_report.html). Multiple runs on the same day each get their own file — no overwriting. Includes:
- Run summary dashboard (trend research results, QA stats)
- Idea scoring distribution + rejection reasons
- Per-project cards (competitive analysis, QA attempts, errors fixed)
- Full retrospective
overview.html
Append-only — existing content is never replaced. Each run injects new project cards and a history entry into well-known HTML markers (<!-- PROJECT_CARDS_START/END -->, <!-- TIMELINE_START/END -->, <!-- STATS_TOTAL -->, <!-- STATS_AVG_SCORE -->). Statistics are recalculated as a running total. Created fresh if it doesn't exist yet. Includes:
- Run history timeline (links to each report)
- Full project catalog with platform / type / score filters
- Statistics dashboard (total projects, stack distribution, average QA attempts, average idea score)
projects/{slug}/user-suggest.html
Generated immediately after each project completes — one standalone file per project. Includes:
- Quick wins, feature enhancements, tech improvements, growth strategies, monetization ideas
- Priority and estimated effort labels per suggestion
Prerequisites
ghCLI authenticated (gh auth login)- Node.js 18+ (for Web / CLI projects)
- Git configured
- Context7 MCP server enabled (for latest stack research)
Fallback Strategy
Four-tier automatic fallback on errors:
1st build-error-resolver agent (auto-fix)
2nd build-error-resolver retry (with scope-reduction hint)
3rd Strip Nice-to-have → re-implement Must-have only → 1 final QA
4th Mark SKIP + log error details → proceed to next project
GitHub push failures are retried up to 3 times with exponential backoff (5s → 10s → 20s).
License
MIT