AI Conversation Auditor & Fixer

A two-skill bundle for prompt engineers and LLM developers. Diagnose, then ship. No vague ratings, no "have you
tried X" — concrete rewrites you can paste into production.

What's in the bundle

🔍 ai-convo-critic — the diagnostician

Forensic critique of AI conversation transcripts and tool-execution logs across three dimensions:

• Prompt Interpretation — did the model parse intent correctly, or invent a task?
• Response Quality — accuracy, completeness, format, reasoning depth
• Tool / Function Calls — right tool, right args, no redundant calls, correct result handling

Output: ranked issue breakdown with severity labels (🔴 Critical → 🔵 Suggestion), each with WHERE / WHAT HAPPENED / WHY IT MATTERS / FIX.

🛠 ai-prompt-engineer — the fixer

Takes critique findings (or just a "this prompt isn't working" complaint) and produces complete, deployment-ready
artifacts — not patches you have to mentally merge:

• Rewritten user prompts
• Full revised system prompts
• Corrected tool / function schemas
• Restructured conversation flows
• Parameterized reusable templates

Includes a hardening checklist for the cold-start case (no prior critique, just "make this more robust").

Why bundle them

They're explicitly designed to pair. The critic is the spec; the engineer is the executor. Use them together for the
natural workflow:

[broken conversation] → ai-convo-critic → [diagnosed issues]
↓
ai-prompt-engineer → [fixed artifact]
↓
[deploy & re-test]

Or use either independently — the engineer runs a lightweight diagnosis itself if invoked without a prior critique.

When to trigger

Either skill auto-fires on natural prompts:

• "Critique this conversation" / "why did the model do X" / "audit this tool call"
• "Fix this prompt" / "rewrite this system prompt" / "the agent keeps doing Y, fix it"
• "Compare response A and response B"
  Either skill auto-fires on natural prompts:
• "Critique this conversation" / "why did the model do X" / "audit this tool call"
• "Fix this prompt" / "rewrite this system prompt" / "the agent keeps doing Y, fix it"
• "Compare response A and response B"
• "Turn this rough prompt into a production-ready one"

What you get vs. what you don't

✅ You get: Specific turn-by-turn diagnosis. Complete rewritten artifacts. Severity-ranked issue lists. Diff annotations and
risk flags on every fix.

❌ You don't get: Generic "your prompt could be clearer" feedback. Numerical 1–10 scores. Half-rewrites with [... rest unchanged ...] placeholders.

Format support

Handles raw chat (User: / Assistant:), OpenAI/Anthropic JSON API format, partial transcripts, multi-agent logs, and
before/after diff pairs.

Host compatibility

Built for Claude Code. Compatible with any host that supports the standard SKILL.md skill format (Codex CLI, OpenClaw, Hermes).