Internal: compare claims within the named document set. Respect "only docs/no code". Internal agreement proves consistency, not runtime truth.
Implementation/contract: compare current-behavior claims with code/config and intended behavior with approved specs/contracts. Inspect related internal contradictions as useful.
Combined: use both when the user asks for both or a general accuracy audit has both sources available. State the chosen scope; if code is unavailable, mark claims unverified instead of implying they match.
Audit only: deliver findings without editing.
Check and fix: when requested, apply evidence-backed repairs to the named local documents, then recheck affected claims and links. Do not silently change implementation to make docs true. Escalate unresolved product/authority choices; continue independent clear repairs.

These are internal modes, not questions the user must answer before work starts.

Workflow

Identify the named document surface, audience, scenarios, and requested action. Do not expand named targets into sibling documents without reason or authorization. With no named targets, inspect README, setup/API docs, and examples relevant to the question.
Build a claim map: behavior, commands, defaults, lifecycle, permissions, terminology, assumptions, acceptance, and cross-links. Record source precedence per class of claim.
For internal evidence, read references/internal-audit.md; distinguish explicit supersession from unsupported claims. Absence of support alone does not prove a claim false or obsolete.
For implementation evidence, read references/checklist.md. Trace documented scenarios to implementation and inspect public/config contracts back toward documentation for important omissions. Cite both locations and identify generated versus canonical sources.
Classify each finding as document defect, implementation gap, internal conflict, or needs evidence/decision. If a doc describes approved future behavior, mark the implementation gap; do not rewrite it to bless current behavior. Consult references/worked-examples.md when authority is disputed.
Recommend update, move, merge, remove, split, or resolve. Separate a suggested repair from an applied edit. Preserve real contracts and unresolved claims explicitly.
If repairs were requested, make the bounded authorized edits and verify them. Route a wholesale rewrite to hai-rewrite-doc with the established evidence; product decisions belong to hai-prd, plan changes to hai-goal.
Read references/output-template.md and report scope, findings, actual edits, verification, and unresolved decisions without repeating a full report for each evidence mode.

Impact

P0: evidenced severe wrong decision, data/security harm, or launch failure.
P1: core scenario, scope, or acceptance is materially wrong.
P2: bounded drift, omission, duplication, or misleading terminology.
P3: minor wording or link issue.
Needs evidence/decision: unresolved; not a confirmed defect.

Rate actual impact, not the presence of a sensitive keyword. A security-related typo alone does not become a release blocker.

Boundaries

Formatting without factual changes → readme-beautifier.
Whole-document reconstruction → hai-rewrite-doc.
Product requirements or acceptance definition → hai-prd.
Entity fields and storage/migration placement → entity-model-auditor.
Proving a software change works → write-technical-acceptance-report; documentation inspection alone does not prove runtime acceptance.

hai-audit-docs

Hai Audit Docs

Principle

Select evidence and action