Vibe Tuning — AI skill for Claude Code

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