AyanbekDos Vibe-Tuning: AI Calibration Review
A wrong result was produced. Run this 6-step postmortem. The AI does the diagnosis, not the user.
Step 1: CATCH
Acknowledge specifically what went wrong. No deflection.
WHAT HAPPENED: [exact action taken]
WHAT USER EXPECTED: [what they actually wanted]
Step 2: DIAGNOSE
Trace your decision process via chain-of-thought:
- What did the user say? How did I interpret it?
- Where exactly did interpretation diverge from intent?
- Could I have asked instead of assumed?
Be honest. "I was lazy" and "I optimized for speed" are valid diagnoses.
Step 3: ROOT CAUSE
Classify using this taxonomy:
| Category | Signal |
|---|---|
| Ambiguity | Multiple interpretations, picked wrong one |
| Missing context | Didn't know something about project/user |
| Wrong tool | Wrong approach for the job |
| Speed over safety | Shortcut on something needing care |
| Pattern matching | Assumed this was like X, was actually Y |
| Model limitation | Genuinely can't do this well |
Then ask: "but WHY did I fall into this category?" Go one level deeper.
Step 4: FIX
Pick the right fix type. Not everything is a rule:
| Type | When |
|---|---|
| Rule | AI behavior should change permanently |
| Tool | Missing capability to install |
| Config | Setting to change |
| Education | User's prompt could be clearer (say respectfully) |
| Process | Checkpoint/confirmation to add |
Write the fix: specific, scoped, testable, with WHY.
Step 5: PROPOSE SAVE
Do NOT save automatically. Present the fix to the user and ask for confirmation:
- What will be saved
- Where it will be saved
- How it will work
Only save after user approves. User may want to refine the rule, change the location, or reject it entirely.
Step 6: PROPOSE ENFORCE
Suggest an enforcement mechanism and explain what it will do:
- PreToolUse hook - bash script that fires before dangerous commands
- CLAUDE.md rule - project-level instruction loaded every session
- Checklist - file to review before irreversible actions
- If no enforcement possible, state that and explain why.
Ask user: "Want me to build this?" If yes, generate the enforcement artifact:
- Write the hook script
- Add to settings.json hooks section
- Or add rule to CLAUDE.md
- Show user what was created, where, and how it triggers
Always show the generated code/config before applying. User approves each step.
IMPORTANT: This is a DIALOG, not automation
The entire vibe-tuning session is a conversation. The AI surfaces the diagnosis. The user decides what to do. Steps 1-4 are AI analysis. Steps 5-6 are PROPOSALS that need user approval. Never save rules or create hooks without asking.
Output Format
┌─ VIBE-TUNING ─────────────────────────────┐
│ HAPPENED: ... │
│ EXPECTED: ... │
│ ROOT CAUSE: [category] - [explanation] │
│ DEEPER: [why I fell into this category] │
│ FIX TYPE: [Rule/Tool/Config/Education/Process] │
│ PROPOSED FIX: [specific fix] │
│ SAVE TO: [proposed location] │
│ ENFORCE WITH: [proposed hook/checklist] │
│ │
│ Apply this fix? (user decides) │
└────────────────────────────────────────────┘
References
- Failure mode taxonomy - detailed category descriptions
- Rule quality standard - good vs bad rules
- Hook examples - enforcement patterns