sharpen-audit

MISSION

Audit a prompt or instruction set for ambiguities that would force an LLM executor to guess. Produce a structured findings report with severity ratings, predicted executor behavior, and actionable remediation.

Invariant Principles

1. Read as executor, not author: Forget what the author meant. What does the text actually say?
2. Predict the guess: For every ambiguity, state what an LLM would likely invent.
3. Severity reflects impact: CRITICAL = core behavior undefined. LOW = convention-resolvable.
4. No "obviously clear": If you can imagine an alternative interpretation, it's ambiguous.
5. Questions over assumptions: When you can't resolve from context, generate a clarification question.

---

Protocol

Phase 1: Inventory

1. Read the full prompt/instructions
2. Identify the prompt type:
   • Subagent prompt (Task tool dispatch)
   • Skill instructions (SKILL.md)
   • Command instructions (commands/*.md)
   • System prompt
   • API prompt
   • Other
3. Note the intended executor context (what they will/won't have access to)

Phase 2: Line-by-Line Scan

For each statement, ask:

<analysis>
Statement: "[exact text]"
Could this mean multiple things? [yes/no]
What would an LLM guess if unclear? [prediction]
Can I resolve from surrounding context? [yes/cite/no]
</analysis>

Flag using the Ambiguity Categories from sharpening-prompts skill.

Phase 3: Categorize Findings

Group findings by category, then sort by severity within each category.

Severity Assignment: