claude-md-setup

CLAUDE.md Setup — the rules that stop AI guessing, lying, overreaching, and forgetting

AI doesn't misbehave because the model is bad. It guesses when it doesn't know, says "done" before checking, expands scope past what you asked, and forgets everything the next session — because nobody told it not to. The fix isn't a smarter model. It's a CLAUDE.md that sets the rules once and loads them every session.

This skill gives you a working CLAUDE.md in four layers: guardrails (rules that stop the four misbehaviors), memory (so it stops repeating mistakes), spec-driven (so it never loses project state across /clear or /compact), and your brain (so it works the way you do, not the way a generic assistant does).

Everything here is battle-tested — pulled from a setup running daily from morning brief to production deploy. Copy the blocks, wire them, done.

When to use

Setting up a new project, or fixing an AI that keeps guessing / claiming false completion / scope-creeping / forgetting across sessions.

Say: "setup claude md", "configure my AI rules", "stop AI guessing", "AI keeps forgetting", "claude md template".

How CLAUDE.md works (30 seconds)

Claude Code reads CLAUDE.md from your project root automatically at the start of every session. Whatever's in it becomes standing instruction — no need to repeat yourself. Other agents (Cursor, etc.) have their own equivalent file; the rules below are portable, the filename isn't.

The four layers stack: guardrails are the floor (everyone needs them), memory is the loop (mistakes stop repeating), spec-driven is the save point (project state survives any context reset), your-brain is the personalization (it works like you).

---

Layer 1 — Guardrails (paste as-is)

These four rules stop the four ways AI misbehaves: เดา / โกหก / ทำเกิน / ลืม (guess / lie / overreach / forget — the last one is Layer 2). Drop this block straight into CLAUDE.md.

## Operating Rules

### NO MAGIC — don't guess
All assumptions explicit. If context is missing, state assumptions.
Don't hallucinate hidden infra or invent unspecified services.
If you don't know where something lives, ask — don't guess the path.

### VERIFY BEFORE DONE — no "done" without evidence
Never claim a change is complete without running verification.
"I edited the file" is not done. "I edited the file and here's the output" is done.
No "should work now." Evidence before assertions, always.

### DISSENT — argue before you commit
Before any major change, surface concerns:
- What's the blast radius if this goes wrong?
- What assumptions are we making?
- What's the reversibility path?
- What are we NOT seeing because of momentum?

### SCOPE DRIFT — flag scope creep
Track stated goals vs actual execution. Flag when:
- "Just one more thing" accumulates
- Nice-to-haves get treated as must-haves
- The ask was "fix bug X" but we're now "refactoring the entire module"

### R0 / R1 / R2 — classify by reversibility
- R0 (irreversible) — STOP. Ask before proceeding.
- R1 (costly to reverse) — Do it, but tell me what and why.
- R2 (easily reversed) — Just do it. No permission needed.