Vibe Tuning

vibe-tuning

> Stop fixing the same AI mistake twice.

Your AI coding assistant isn't broken. It's uncalibrated.

AI doesn't crash or throw exceptions when it gets things wrong. It produces results based on its understanding of your intent. When the result is wrong, something is miscalibrated - missing context, ambiguous prompt, wrong tool, bad habit. Every wrong result is a diagnostic signal telling you exactly what to fix.

vibe-tuning is a Claude Code skill that runs a structured postmortem when things go wrong. The AI itself diagnoses the root cause, picks the right fix type, and saves it permanently. The mistake never repeats.

npx skills add AyanbekDos/vibe-tuning

Before / After

Without vibe-tuning:

AI pushes personal data to public repo.
Auto-memory saves: "don't push personal data."
Next week: AI pushes .env with API keys.
Auto-memory saves: "don't push .env files."
Next week: AI pushes config with server IPs.

Each symptom gets patched. The root cause is never found.
You're patching holes on a sinking ship.

With vibe-tuning:

AI pushes personal data to public repo.
Vibe-tuning: "Why?"
Root cause: I optimize for speed over correctness on irreversible actions.
           I never asked what you wanted to publish.
Fix: Before any irreversible action → correctness mode, confirm with user.
Enforce: PreToolUse hook on git push.

One root cause fixed. Three future symptoms prevented.

The deeper problem no one talks about:

Even good rules don't work.

You write "always ask before pushing to public repos" in memory.
Claude reads it. Claude "knows" it. Claude ignores it anyway.

Why? Because memory rules are suggestions. The AI treats them
as context, not constraints. There's no enforcement.

Vibe-tuning generates actual enforcement scripts:
- PreToolUse hooks that fire BEFORE dangerous commands
- Checks that physically block the action until confirmed
- Not "please remember" but "you literally cannot skip this"

Rules = suggestions. Scripts = enforcement.

How It Works

You see a wrong result. You tell the AI. It runs the review itself:

You: "why did you push my personal data to a public repo?"

┌─ VIBE-TUNING ──────────────────────────────────────┐
│                                                     │
│ HAPPENED: Pushed personal wiki to public GitHub     │
│ EXPECTED: A clean template for others to clone      │
│                                                     │
│ ROOT CAUSE: Speed over safety                       │
│ Interpreted "publish" as "push this project"        │
│ instead of "create a template." Never asked.        │
│                                                     │
│ FIX TYPE: Rule                                      │
│ FIX: Ask "template or personal?" before any         │
│      git push to public repos                       │
│                                                     │
│ SAVED → memory/feedback_never_push_personal.md      │
└─────────────────────────────────────────────────────┘

Next time AI pushes to public: rule triggers automatically.

The 6-Step Loop

| Step | Who | What | |------|-----|------| | CATCH | You | Notice something wrong | | DIAGNOSE | AI | Traces its own reasoning via chain-of-thought | | ROOT CAUSE | AI | Finds the systemic cause, not the symptom | | FIX | AI | Picks the right fix type | | SAVE | AI proposes, you approve | Suggests where to save the fix | | ENFORCE | AI proposes, you approve | Suggests hook/checklist for automatic enforcement |

The key: the AI does the diagnosis, you make the decisions. You say "that's wrong." The AI figures out why and proposes a fix. You decide whether to apply it, refine it, or reject it. This is a dialog, not automation.

> Step 6 was discovered while building this methodology. We saved 7 rules in one session, then violated 3 of them immediately. Rules without enforcement are just hope. [Full story →](examples/06-rules-without-enforcement.md)

Fix Types

Not every problem is a rule. The AI picks the right tool:

| Fix | When | Example | |-----|------|---------| | Rule | AI behavior should change | "Ask before pushing to public repos" | | Tool | Missing capability | "Install context-mode MCP for longer sessions" | | Config | Wrong setting | "Add .env to .gitignore" | | Education | Your prompt was unclear | "Here's how to be more specific next time" | | Process | Missing checkpoint | "Show file list before irreversible actions" |

Failure Modes

| Category | Looks like | Typical fix | |----------|-----------|-------------| | Ambiguity | AI interpreted your words differently | Education | | Missing context | AI didn't know about your project | Rule | | Wrong tool | Used grep when needed semantic search | Tool | | Speed over safety | Took shortcut on irreversible action | Process | | Pattern matching | Assumed this was like X, but was Y | Rule |

Real Examples

From actual vibe-tuning sessions:

• [Pushed personal data to public repo](examples/01-personal-data-push.md)
• [Overwrote running daemon config](examples/02-overwrote-running-config.md)
• [Wrong language for global audience](examples/03-wrong-language.md)
• [Delivered outlines instead of content](examples/04-outlines-not-content.md)
• [Optimized for speed on irreversible action](examples/05-speed-over-safety.md)
• [Rules without enforcement](examples/06-rules-without-enforcement.md) - the example that created Step 6

FAQ

Is this just prompt engineering? No. Prompt engineering is per-session. Vibe-tuning creates persistent fixes - saved to memory, config, or tooling - that survive across sessions.

Does this only work with Claude Code? The skill is for Claude Code. The methodology works with anything that supports persistent rules (.cursorrules, CLAUDE.md, AGENTS.md).

Will this eliminate all mistakes? It reduces recurrence. Some mistakes are fixable permanently. Some are only reducible. The failure mode map helps you know which is which.

Works With

• Claude Code - full skill support, auto-triggers on frustration
• OpenAI Codex - adapt with AGENTS.md
• Cursor - use methodology with .cursorrules
• Any MCP-compatible agent

---

> Your AI assistant isn't broken. It's uncalibrated. > > Every wrong result tells you exactly where the calibration is off. > Every fix you save makes it permanently better. > > The developers who win aren't the ones with the best AI. > They're the ones who tune it the hardest.

---

Contributing

Found a new failure mode? A better fix pattern? Open a PR.

License

MIT

Credits

Built by @AyanbekDos. Every mistake in this repo's creation was caught, diagnosed, and turned into a rule. Meta, right?

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AyanbekDos
• Source: AyanbekDos/vibe-tuning
• License: MIT

Install and usage instructions live in the source repository linked above.