01
Less code.
Delete, reuse, and simplify before adding more things for someone else to maintain.
For coding agents who need supervision
Keep it simple, stupid.
A practical, token-efficient rulebook for stopping a two-file change from becoming a six-layer architecture.
No package. No custom index. No “enterprise-ready” abstraction for a button. ✨
What it fixes
Less nonsense. More working software.
02
Less guessing.
Start with the files you provided. Search the repository only when the context is actually insufficient.
03
Less AI slop. 🧹
Use the existing design system. Not every product needs purple glows, glass cards, and a three-column feature grid.
04
More proof. 🧪
New API endpoints are tested locally before anyone confidently writes “should work.”
Included
Six files. Calm down.
Enough structure to keep agents disciplined. Not enough structure to require an architecture committee.
Using Claude Code, Gemini CLI, or another coding agent?
You may need to rename AGENTS.md to the instruction filename
expected by your tool (for example, CLAUDE.md or
GEMINI.md).
Learn more
AGENTS.md
agent-guides/
├── engineering.md
├── backend.md
├── frontend.md
├── uiux-ds.md
└── testing.md⚠️ What it does not do
Mostly, it gets out of the way.
- Install another framework.
- Read the whole repository for a button change.
- Create abstractions for imaginary future features.
- Test your API on production. Obviously.
- Replace your judgment. Sadly, you still need that.
That is literally it
Your agent already knows how to code.
It just needs rules telling it when to stop doing too much.