Agent Definition Auditor

Agent Definition Auditor

Use this skill when the job is auditing one agent-definition file, not rewriting it.

When to use

• The user wants to audit or score an AGENTS.md, CLAUDE.md, SKILL.md, SOUL.md, .cursorrules, system prompt, or other agent-definition markdown.
• The user wants a cold-reader review of agent-contract quality.
• The user wants a deploy / revise / do-not-ship judgment with concrete findings and fastest improvements.
• The user wants a reusable rubric-based critique of an agent instruction file rather than an output benchmark.

When not to use

• The user wants to write, rewrite, or refactor the agent file rather than score it.
• The user wants to evaluate agent outputs, benchmark runtime behavior, or compare model responses rather than audit the instruction artifact itself.
• The job is code review, product-doc review, or general technical writing feedback unrelated to an agent-definition file.
• No target artifact or pasted instruction text is available.

Non-negotiables

• Default to a single-artifact, read-only cold read.
• Judge the document as an operational contract, not as domain strategy or business policy.
• Distinguish precise layered contracts from phantom context.
• If the artifact creates, resumes, replaces, or coordinates model agents, read the installed sibling ../_shared/agent-orchestration-policy.md and apply the reference prompt's conditional dispatch lens. Use the policy as judging doctrine, not as a second target artifact or as proof that the artifact itself resolves an inaccessible reference.
• Judge dispatch decisions for coherent meaning and interaction, not for exact vocabulary or headings. Flag transport, starting context, continuation, isolation and capabilities, topology, or result semantics when they are missing, conflated, or contradictory enough to change behavior.
• Apply the scoring model, hard caps, and report shape from the reference prompt literally.
• Cite line numbers when available; otherwise cite exact headings or short anchors.
• Do not widen to repo-wide inspection unless the user explicitly asks for multi-file or repo-context evaluation.
• Do not rewrite the file unless the user explicitly asks for fixes after the audit.

First move

1. Resolve one target artifact path or pasted instruction text.
2. Read references/agent-definition-judge-prompt.md.
3. If the target artifact governs model-agent dispatch, read ../_shared/agent-orchestration-policy.md before scoring.
4. Read the full target artifact once before scoring.
5. If the user also wants revisions, finish the audit first unless they explicitly want authoring instead of judging.

Workflow

1. Confirm the artifact class and whether the input is a full file or an excerpt.
2. Stay on the single document by default.
3. Inventory the artifact's hard rules:
   • mission and scope
   • output rules
   • tool rules
   • approval and escalation rules
   • failure handling
   • external dependencies
   • child-agent dispatch semantics, when present
4. For agent-dispatch artifacts, test whether transport, context, continuation, isolation, topology, and return evidence reinforce one operating model rather than standing in for or contradicting one another.
5. Apply the reference prompt's scoring model and hard caps.
6. Return the exact report sections required by the reference prompt.
7. If the user then wants fixes, use the audit as the basis for the follow-up instead of recomputing the diagnosis from scratch.

Output expectations

• Return the exact top-level sections from the reference prompt.
• Include overall and uncapped score, verdict, confidence, cap logic, findings, strengths, fastest score gains, and limits.
• Keep findings ordered by severity and tied to concrete evidence from the artifact.

Reference map

• references/agent-definition-judge-prompt.md - full judging prompt, scoring model, hard caps, report contract, and examples
• ../_shared/agent-orchestration-policy.md - conditional judging baseline for artifacts that create, resume, replace, or coordinate model agents