sstklen

中文版 · 日本語版

YES.md

PUA says NO. YES says YES.

A Claude Code skill that makes AI safe, evidence-based, and thorough — with encouragement, not pressure.

別 PUA 你的 AI，YES 它。

🌐 English | 中文 | 日本語

Why Your CLAUDE.md Doesn't Work

Your CLAUDE.md has one layer. AI reads it, then chooses whether to follow it.

YES.md has six layers. AI has to get past all six to cut corners.

Layer	What	How	Who Enforces
1	Format — AI actually understands your rules	AI.MD converts prose to structured-label format	Format
2	Trigger — activates at exactly the right moment	20+ specific trigger conditions in SKILL.md description	Skill engine
3	Hooks — blocks dangerous commands, reminds checks	Scripts run automatically before/after AI actions	Machine
4	Anti-Slack — 7 lazy patterns caught and corrected	Deflecting, guessing, spinning, surface-fix, empty questions...	AI self-correction
5	Gates — no backup = no edit, no test = no "done"	Hard stop rules that halt progress until requirements are met	AI self-discipline
6	Memory — mistakes get recorded, never repeated	Error logs + knowledge base persist across sessions	External storage

Layer 3 is the key. Everything else relies on AI self-discipline. Hooks are scripts that run automatically — the AI cannot bypass them even if it wants to.

Most tools give you Layer 4 or 5 alone. That's why they don't work. Six layers together is why YES.md works.

The Problem

AI coding assistants have seven bad habits:

Bad Habit	What Happens
🎲 Guessing	"Probably a permissions issue" — without checking
🙈 Deflecting	"Please check your environment manually"
🩹 Surface fixing	Fixes one bug, ignores three related ones
🔄 Blind retrying	Same command 3 times, then gives up
🤷 Empty questions	"Can you confirm X?" — without investigating X
💬 Advice, not action	"I suggest..." instead of actual code
🛠️ Ignoring tools	Has WebSearch but guesses. Has Bash but doesn't run.

PUA-style skills fix habit #4 (blind retry / giving up).

YES.md fixes all seven.

How It Works

Three Pillars

Pillar	What It Does
Safety Gates	Backup before editing. Check blast radius. Verify after changes. Pre-flight before deploy.
Evidence Rules	No guessing. No "probably." Every claim needs data, line numbers, or tool output.
Ripple Awareness	After every fix: check same pattern elsewhere, check upstream/downstream, check edge cases.

Hooks (Machine-Enforced)

Rules rely on AI self-discipline. Hooks don't. They are scripts that run automatically before and after AI actions.

Hook	When	What It Does
pre-bash-guard.sh	Before any Bash command	Blocks rm -rf, git reset --hard, DROP TABLE, and other destructive commands
post-edit-check.sh	After any Write/Edit	Reminds AI to run ripple-check before saying "done"
post-deploy-health.sh	After deploy commands	Reminds AI to verify containers, endpoints, and logs

The AI cannot skip these. That's the point.

Anti-Slack Detection

YES.md identifies 7 specific AI avoidance behaviors and forces self-correction — without waiting for the user to notice:

Deflecting to user     → Do it yourself first
Unverified blame       → Run the command, then speak
Spinning in circles    → Full stop, fundamentally different approach
Surface-only fix       → Ripple Check
Empty questions        → Investigate first, then ask
Advice without action  → Ship code, not suggestions
Tool neglect           → Use the tool, not your memory

Debugging Escalation

Failures	Action
2	Switch: fundamentally different approach (not parameter tweaks)
3	Five-Step Audit: read error word-by-word → WebSearch → read 50 lines context → verify assumptions → invert hypothesis
4	Isolate: minimal reproduction
5+	Structured Handoff: verified facts, eliminated causes, narrowed scope, next steps

Key difference from PUA: Level 3 forces a direction check BEFORE continuing. Persistence in the wrong direction is worse than stopping.

Conclusion Integrity

Before any root-cause claim, four mandatory questions:

1. Data source? — log / DB / API / curl?
2. Time range? — all data or just recent?
3. Sample vs total? — how much did you see?
4. Other possibilities? — what else could explain this?

Incomplete data = "⚠️ Based on partial data..." (never "definitely" or "the culprit is").

YES.md vs PUA