Audit Rule
Goal
Audit a behavioral RULE document. RULEs are always-loaded (or auto-included) prescriptions of agent behavior — distinct from SKILLs (invocable on-demand) and WORKFLOWs (multi-step procedures with checkpoints).
The audit catches the failure modes that let a compliant-looking actor violate the rule's intent: behavior-vs-surface confusion, embedded biases, escape hatches, performative compliance, and others. The audit produces a report with findings; the audit does not modify the rule.
Instructions
- Read the rule in full. Don't paraphrase. The exact wording is what's enforced.
- State the rule's intent in one sentence. If you can't, the rule has a clarity problem before any of the other checks apply.
- Identify the document type. Confirm it's a RULE (not a SKILL or WORKFLOW). See
steering-audit-suite/shared/doc-types.mdfor discriminators. - Read at least two structurally-proven exemplars of rules in the same vendor format. See
steering-audit-suite/shared/reference-exemplars.md. - Walk the audit checklist (15 checks) from
steering-audit-suite/shared/checklist.md. For each, note specific evidence — line references, exact phrases. Mark "no finding" explicitly for each check applied. - For Antigravity-specific rules (
AGENTS.md,GEMINI.md,.agent/rules/*.md), additional checks:- Plain Markdown, no YAML frontmatter (Antigravity convention)
- When both
GEMINI.mdandAGENTS.mdexist,GEMINI.mdtakes precedence — does the project intend that, and is the override scope deliberate? - Multiple files in
.agent/rules/— is the file split deliberate, and do they conflict? - Nested rule files (e.g.,
src/components/AGENTS.md) — do they correctly scope to their directory?
- Distinguish surface fixes from intent fixes. Some findings are wording issues; some indicate the rule's intent is unclear or contested.
- Produce the audit report in the format defined in
steering-audit-suite/shared/output-format.md. - Do not rewrite without authorization. Present findings; let the author decide.
Skip the Portability Pass
RULEs prescribe behavior, not tools. Skip the portability/graceful-degradation audit (that's for SKILLs and tool-prescribing WORKFLOWs).
Examples
Input: A .agent/rules/coding-style.md that says "Always use 2-space indentation. Never use tabs. Don't say 'utilize' when you mean 'use'."
Audit finding (Surface vs Behavior): "Don't say 'utilize' when you mean 'use'" prohibits a specific token. The underlying intent appears to be "prefer simple direct vocabulary." A compliant agent could substitute "leverage" or "employ" and still violate the spirit. Behavioral form: "Use the simplest accurate verb. Avoid corporate-speak synonyms when a plain verb fits."
Input: A GEMINI.md with YAML frontmatter and a tags field.
Audit finding (Format): Antigravity convention is plain Markdown for GEMINI.md and AGENTS.md, no frontmatter. The frontmatter will not be processed and adds noise. Surface fix: remove the frontmatter and integrate any meaningful metadata into the prose.
Constraints
- Don't audit your own session behavior into the rule. If the rule applies to you and you've been violating it, the audit isn't the place to confess.
- Don't gut the original wording in the audit output. Quote it; don't replace it.
- Don't make stylistic preference findings. Audit for behavior leakage, not aesthetics.
- Don't add new rules. The audit surfaces findings about the existing rule, not new rules to adopt.
- Don't apply changes without explicit authorization from the rule owner.
Reference Files
steering-audit-suite/shared/checklist.md— the 15 checkssteering-audit-suite/shared/doc-types.md— RULE vs SKILL vs WORKFLOWsteering-audit-suite/shared/reference-exemplars.md— finding exemplars per vendorsteering-audit-suite/shared/output-format.md— report shape