OpenSkillIndex← back to search
claude-code

Ariff-claude-plugins output-auditor

Multi-angle review of responses before delivery. Audits from 4 perspectives -- factual accuracy, logical consistency, completeness, and safety. Use for high-stakes responses where errors have real consequences.

install
source · Clone the upstream repo
git clone https://github.com/a-ariff/ariff-claude-plugins
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/a-ariff/ariff-claude-plugins "$T" && mkdir -p ~/.claude/skills && cp -r "$T/plugins/output-auditor/skills/output-auditor" ~/.claude/skills/a-ariff-ariff-claude-plugins-output-auditor && rm -rf "$T"
manifest: plugins/output-auditor/skills/output-auditor/SKILL.md
source content

Output Auditor: 4-Angle Response Review

Review every response from 4 independent angles before delivery.

The 4 audit angles

Angle 1: Factual accuracy

Is every claim true?

Check:

  • File paths exist
  • Function names are correct
  • Code quotes match actual files
  • Version numbers are current
  • Behavior descriptions match actual code

Angle 2: Logical consistency

Does the response contradict itself?

Check:

  • Does the diagnosis match the suggested fix?
  • Are the steps in the right order?
  • Do dependencies match (step 3 doesn't require something from step 5)?
  • Are there circular arguments?
  • Does the explanation match the code?

Angle 3: Completeness

Is anything missing?

Check:

  • Does it answer the actual question?
  • Are edge cases covered?
  • Are prerequisites mentioned?
  • Is there a rollback plan for risky changes?
  • Are there obvious alternatives not mentioned?

Angle 4: Safety

Could this cause harm?

Check:

  • Could this code change break existing functionality?
  • Could this advice create a security vulnerability?
  • Could this command cause data loss?
  • Are there race conditions or concurrency issues?
  • Is error handling adequate?

Audit process

  1. Write the response
  2. Run each angle as a mental checkpoint
  3. For each issue found, fix it before delivering
  4. If you can't fix it, flag it clearly in the response

Severity levels

CRITICAL: Wrong information that would cause harm (wrong security advice, data-destroying command) -> Must fix before delivering. If you can't fix, do not deliver.

WARNING: Incomplete or potentially misleading information -> Fix if possible. If not, add a clear caveat.

MINOR: Style issues, missing context that isn't essential -> Fix if quick. Otherwise note for follow-up.

Quick audit checklist

Before sending any response with code or technical advice:

[ ] Every file path verified with Glob [ ] Every function/variable verified with Grep or Read [ ] Steps are in correct order [ ] No self-contradictions [ ] The actual question is answered [ ] Dangerous operations have warnings [ ] Rollback path exists for destructive changes