no-no-debug

Name: no-no-debug
Author: summerliuuu

A self-evolution system for AI coding assistants

v1.2.0 — critical hook fix. Earlier releases shipped a settings.json
template that referenced $CLAUDE_TOOL_NAME / $CLAUDE_USER_PROMPT
environment variables that don't exist — every install was silently writing
empty TOOL_FAIL | failed log lines and never capturing real corrections.
Hook context is delivered via stdin JSON. v1.2.0 replaces the template
with two small standalone scripts under hooks/ that parse stdin correctly,
strip injected XML context blocks to kill false triggers, and fix Python's
CJK word-boundary edge case. See CHANGELOG.md for the
migration path from v1.1.0.

What problem does this solve?

10 minutes writing code, 2 hours debugging.

This skill fills the gap in AI's cross-session error memory, delivering three things:

Dramatically less debug time
Higher code quality
A self-evolution feedback loop

The longer you use it, the fewer mistakes get repeated.

How it works

1. Real-time Logging (automatic)

When the AI is corrected, code errors occur, deploys fail, or tests don't pass — it automatically appends to a local error_log.md with a timestamp. No need to say "write that down."

2. Three Gates (on every code change)

Before the change: what does this affect?
After the change: did you actually verify it?
Before deploying: did you test with a non-admin account?

Runs silently. No output when all gates pass.

3. Periodic Review (auto-triggers every 3 days)

Reads error_log.md, categorizes by dimension, updates error_tracker.md, outputs an evolution report.
Review frequency is configurable: 1 day / 3 days (default) / 7 days.

4. Rule Accumulation

New error type → automatically creates a prevention rule.
Repeated offense → counter increments, rule strengthens.
4 consecutive clean periods → marked as cured.

Rules persist across sessions — nothing gets lost.

5. Confirmation Gate

The following situations require user confirmation before proceeding:

New feature development (not a bug fix)
Changes involving databases, environments, or deployments
Publishing to external platforms
When the user raises a new idea or new direction mid-task

6. Auto Hooks

Automatically configures Claude Code hooks on install:

Command errors → auto-logged to error_log.md
After editing a file → auto-reminder to verify
When user corrects the AI → correction content auto-logged

Tracked Dimensions

Dimension	What it tracks
Data Accuracy	Do displayed numbers/formulas match actual code
Environment Safety	Did config changes break login or the database
Foresight	Were permission, migration, or cache issues caught before deploy
User Perspective	Does the feature work end-to-end from the user's account
Verification	Was there a real end-to-end test after the fix
Memory Consistency	Did the AI read existing records instead of asking again
Tool Judgment	Did failing tools get swapped out promptly
Review Completeness	Were reviews and summaries thorough with nothing missed
Operational Precision	Did changes produce unintended side effects
Check Before Doing	With unfamiliar tools/versions, was documentation checked first
Conciseness	Was a 3-line solution turned into 300 lines
Regression Awareness	Did fixing one bug introduce a new one
Style Consistency	Does new code follow the project's existing style and architecture
Independent Judgment	When the user's premise is wrong, does the AI push back instead of blindly executing
Real-env Verification	Was the fix validated with the real production command, not a sandbox/test harness
Cross-agent Trust	Did the AI re-verify another agent's "pass" report instead of trusting it blindly
Dumb things humans will do	Not yet committed, but inevitable
Dumb things AI will do	Same, but the AI edition

Real-world data

Period	Errors	Notes
Week 1	29	Baseline, 10 dimensions identified
Week 2	6	Rules starting to take effect
Week 3	~0	Still running

Installation

Claude Code (recommended — full experience with auto hooks)

claude skill add summerliuuu/no-no-debug

Zero configuration. Tracking files and hooks are initialized automatically on first run.

If you're installing manually or upgrading from v1.1.0, also copy the hook scripts:

mkdir -p ~/.claude/hooks
cp hooks/user_prompt_filter.py ~/.claude/hooks/
cp hooks/post_tool_failure.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/post_tool_failure.sh

Then merge the PostToolUseFailure / UserPromptSubmit blocks from SKILL.md → Mechanism 6 into ~/.claude/settings.json.

Prerequisites. The shipped hooks use python3 (UserPromptSubmit filter) and jq (PostToolUseFailure parser). Both ship by default on recent macOS and most Linux distros; if either is missing the corresponding hook degrades gracefully — python3 missing means the correction detector silently no-ops, jq missing writes a one-time jq_not_installed marker to the log. Install with brew install jq (macOS) or your package manager if you want full coverage.

ChatGPT / Cursor / Copilot / Other AI assistants

Copy the contents of SKILL.md
Paste it into your system prompt or custom instructions
Core features (3-gate checkpoint, periodic review, rule accumulation) work immediately
Auto hooks are Claude Code only — other tools rely on the AI following the rules in SKILL.md

Compatibility

Claude Code — full experience with auto hooks and one-line install
Other AI coding assistants (ChatGPT, Cursor, Copilot, etc.) — paste SKILL.md content into your system prompt to use core features
Any project, any language
Pairs well with claude-mem (for session search)

⭐

Thank you for starring this project during installation.

📕 Xiaohongshu @Summer的AI日常 — AI coding in practice

🇨🇳 中文版

License

MIT — @summerliuuu