AI's biggest enemy is itself — it forgets the bugs it fixed yesterday, falls into the same traps over and over.
6 self-evolving Skills that install code discipline into AI agents. Not about writing faster — about writing code that stays fixed.
If you code with AI, you will hit all 5 of these pain points:
| # | Pain | What Happens |
|---|---|---|
| 1 | 🧠 AI Amnesia | Bug fixed yesterday? Back today. Fix module B, silently break module A. You won't know until users complain. |
| 2 | 🔁 Repeat Traps | Same mistakes — || eats 0, <select> missing placeholder — repeated 5 times. AI has no "learn from pain" ability. |
| 3 | 📝 No Trace | No changelog. What did I touch? What did I break? No starting point for debugging. |
| 4 | 🧪 No Regression | One CSS line → whole layout broken. Manual regression testing is exhausting. Skipping it is terrifying. |
| 5 | 🗺️ Lost in Code | Open a file — what's coupled to it? How was it last modified? Where are the landmines? Every edit is an expedition. |
Vibe Coding Rules turns all 5 into automated workflows. You don't need to know how to code. The AI remembers, checks, and tests for you.
You say: "Fix the save button on the settings page"
Here's what happens step by step:
The AI doesn't jump straight into coding. It pauses:
- Opens
settings.jsand finds the 🍞 breadcrumb header:@COUPLED: server/routes/settings.js, public/js/validation.js @BUGFIX 2026-06-20: Form field name mismatch causing silent save failure - The AI instantly knows: don't just fix the frontend — check the backend route and validation too. And don't revert the field name fix from last time.
Before writing a single line, the AI already knows the full impact scope.
If API verification is needed, safe-terminal-executor auto-wraps curl/wget commands into Node scripts with timeout protection — no more hanging on network calls.
Code is written, but the AI doesn't say "done" yet. It checks 28 rules:
- ✅
||didn't eat a numeric0? (Trap A) - ✅
<select>has a placeholder option? (Trap B) - ✅ POST upsert doesn't hardcode fields? (Trap C)
- ✅ Dual-entry HTML
<script>tags synced on both sides? - ✅ 🍞 breadcrumbs updated?
- ... 28 rules total, run automatically.
Code isn't "good enough." It passes 28 rules.
Code is written, rules are checked — but the AI opens a browser and verifies:
- DOM assertion: Is the save button there? Is it clickable?
- Screenshot diff: Did the layout break vs. last screenshot?
- OCR check: Is the text on the page correct?
One CSS line won't silently destroy your layout. The AI checks with its own eyes.
Everything passes? Auto-record:
- Which files changed
- Lines added / deleted
- 28-rule pass rate
- Any new rules promoted (from observation to hard rule)
coding-principles (pre-code check)
→ write code (safe execution)
→ self-check (28 rules)
→ web-testing (browser verification)
→ changelog (change record)
→ ✅ delivered
You say one sentence. The AI handles the rest in a closed loop.
| Skill | What It Does |
|---|---|
| 🥇 coding-principles | Pre-code: 5 principles + 🍞 breadcrumb pre-scan + 3 auto-trap detection |
| 🛡️ safe-terminal-executor | Safe terminal: all network commands wrapped with timeout + forced exit |
| 🔍 self-check | Post-code: 28 rules + integrity gate + 🍞 breadcrumb seeding + trust-tiered growth |
| 🧪 web-testing | Auto-test: DOM assertion + screenshot diff + OCR verification |
| 📝 changelog | Auto changelog: what changed + rule growth detection |
| 🚀 pipeline-init | One-click bootstrap: conversational Q&A → complete pipeline in 1 minute |
This is the core mechanism. Every time code changes, the AI plants breadcrumbs in file headers:
/*
* ─── 🍞 AI Breadcrumb Navigation ────────────────────
* Tag legend: @COUPLED=coupled file @GOTCHA=trap @BUGFIX=bug fix @MAGIC=magic number
* @DEPENDS=external dependency @ASSUME=assumption @TODO=todo @WHY=design rationale
* @PERF=performance @CONTRACT=interface contract 📖=documentation reference
*
* Breadcrumbs (what depends on this):
* @COUPLED sync field changes with: public/js/chat-core.js#L200, server/routes/chat.js
* 📖 see: docs/architecture/chat-system.md
* @BUGFIX 2026-07-01: mobile overflow layout collapse, fix: add overflow-x:hidden to container
* @GOTCHA: state machine reset in openTab(), changing order causes double-init
* ────────────────────────────────────────────────
*/Pre-scan: Open a file → read 🍞 first → know couplings & past pitfalls → don't repeat mistakes
Post-seed: Auto-update 🍞 after changes → next AI that opens this file has a complete map
Not every change needs all 6 Skills. A 3-line CSS fix ≠ a new feature.
| Level | Trigger | Skills | Use Case |
|---|---|---|---|
| L1 · Light | ≤ 3 files, ≤ 30 lines changed | 5 principles check | Bug fixes, text edits, CSS tweaks |
| L2 · Standard | 3-10 files, or touches routing/data flow | L1 + 28-rule self-check | Cross-module bugs, medium changes |
| L3 · Full | > 10 files, or new feature/refactor | L1 + L2 + auto-test + changelog | New features, architecture changes |
No context waste. A 3-line CSS fix doesn't trigger the full test suite.
# CodeBuddy / Cursor / Windsurf
openclaw skills install @ron-dali/vibe-coding-rules
# Or manual clone
git clone https://github.com/Ron-dali/vibe-coding-rules.gitAfter install, tell your AI: "Initialize pipeline" — it auto-detects your project and configures everything.
Start using in 30 seconds. Zero programming knowledge required.
| If You | Why You Need This |
|---|---|
| 🐣 AI-coder, non-programmer | Your code-discipline exoskeleton — AI stops reckless edits, auto-verifies after |
| 💻 Solo dev / indie maker | You ARE your QA — these 6 Skills ARE your QA team |
| 🏢 Small team + AI | Unified quality standard, less code review overhead |
| 🤖 Heavy AI agent user | Agent gets smarter over time — rules grow, nothing forgotten |
┌─────────────────────────────────────────┐
│ User says one thing │
│ "Fix the save button" │
└──────────────────┬──────────────────────┘
│
┌────────────▼────────────┐
│ 🥇 coding-principles │
│ Pre: 5 principles + 🍞 │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ 🛡️ safe-terminal │
│ Safe network commands │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ ✏️ Write Code │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ 🔍 self-check │
│ Post: 28 rules + gate │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ 🧪 web-testing │
│ DOM + screenshot + OCR │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ 📝 changelog │
│ Changelog + growth │
└────────────┬────────────┘
│
✅ Closed Loop
| Feature | Description |
|---|---|
| 🍞 Breadcrumb Memory | Auto-record couplings + historical pitfalls in file headers — AI never forgets |
| 🔄 Self-Growing Rules | Repeated pitfalls auto-promote to hard rules; false positives auto-retire |
| 🎚️ Progressive Adoption | L1/L2/L3 matched to change size — no context waste |
| 🛑 Integrity Gate | Critical violations block immediately — problems don't propagate |
| 🌍 Cross-Platform | Works with CodeBuddy, Cursor, Windsurf |
| 🔄 Failure Recovery | Multi-violation grouped recovery, delivery unblocked |
| 🔌 Template-Free | Pure prompt + pipeline.json, no framework lock-in |
I don't know how to code. Not a single line.
45 days. Pure AI coding tools. Built a 140,000-line, 3-platform (Web + Desktop + MiniApp) product from scratch — TangBuPing, an AI knowledge democracy platform.
The biggest pain wasn't AI failing to write code. It was AI forgetting what it wrote. Bug fixed yesterday? Back today. Module A fixed, Module B destroyed.
3 months of battle scars, distilled into these 6 Skills. Open-sourced so you don't have to bleed the same way.
We don't hustle harder. We make it so others can't stay lying down either.
Apache 2.0 — Free to use, modify, distribute.