How It Works

Before:

Session start loads:
├── All docs           → 8,000 tokens
├── Old sessions       → 2,000 tokens
├── Task history       → 1,000 tokens
└── Misc               → ???
                Total: ~11,000 tokens

After:

Session start loads:
├── CLAUDE.md          → 450 tokens
├── COMMON_MISTAKES.md → 350 tokens
├── QUICK_START.md     → 100 tokens
└── ARCHITECTURE_MAP.md→ 150 tokens
                Total: ~800 tokens (90% reduction)

Everything else?
Available when you ask, but costs 0 tokens until you do.

Framework-Specific Setup

Got patterns for 9 frameworks with common mistakes and best practices:

| Framework | Example | |-----------|---------| | Express.js | examples/express.md | | Next.js | examples/nextjs.md | | Vue.js | examples/vue.md | | Nuxt.js | examples/nuxtjs.md | | Angular | examples/angular.md | | Django | examples/django.md | | Rails | examples/rails.md | | NestJS | examples/nestjs.md | | Laravel | examples/laravel.md |

Each includes the top 5 critical mistakes for that framework (N+1 queries, auth issues, etc).

Manual Setup

If you want more control or framework-specific patterns:

1. Copy UNIVERSAL_SETUP.md
2. Paste it into Claude Code
3. Answer Claude's questions
4. Get framework-specific patterns included

After Setup

1. Add your critical bugs to .claude/COMMON_MISTAKES.md Only add bugs that took >1 hour to debug. Keep it short.

2. Document your commands in .claude/QUICK_START.md npm run dev, database commands, whatever you use daily.

3. Map your architecture in .claude/ARCHITECTURE_MAP.md Where controllers live, how routing works, etc.

4. Create topic files in docs/learnings/ as you learn One file per topic (testing, API design, deployment)

5. Archive completed work Task docs go to .claude/completions/, old sessions to .claude/sessions/archive/

Claude loads only what it needs. Everything else sits there at 0 token cost until you explicitly ask for it.

Real Numbers

My RedwoodJS project:

• Before: 8,000 tokens at startup, 11,000 total
• After: 800 tokens at startup, 1,300 total
• 7,200 tokens freed up for actual code

Your mileage will vary, but 83-87% reduction is typical across frameworks.

FAQ

Is the bash script safe? It only creates files, never deletes. You can read the source - it's ~470 lines.

Works with my framework? Yes. The bash script creates a universal structure. Works with any language/framework. We have specific patterns for 8 popular ones, but it's not required.

What if I have existing docs? The setup works alongside them. You can migrate gradually or keep both.

Can I customize it? Everything. Edit files, change structure, adjust what loads. It's just markdown files and a .claudeignore.

How do I maintain it? When you hit a critical bug that took >1 hour, add it to COMMON_MISTAKES.md. That's the main upkeep. Rest is optional.

Related Projects

Claude Workflows - Systematic workflow commands for Claude Code

While this repo optimizes what Claude knows (documentation), Claude Workflows optimizes what Claude does (processes). They work great together:

• This repo: Documentation structure (4 core files, 0-token historical context)
• Workflows: Process automation (/dev:start-feature, /git:commit, /test:run)

Both use the same principle: 0 tokens until you need them.

Contributing

Want to add a framework? Copy one of the examples, test it on a real project, and submit a PR with token savings.

Needed:

• Flask/FastAPI
• Spring Boot
• Go (Gin/Echo/Fiber)
• Rust (Actix/Axum)
• Phoenix
• ASP.NET Core

See CONTRIBUTING.md.

Need Help?

• Bug? Open an issue
• Question? Start a discussion
• Love it? Star the repo

License

MIT - use it however you want.

Built because I was tired of Claude loading 11,000 tokens of docs I wrote 3 months ago.

Now it loads 800 tokens and I actually have room to code.