Agents · documentation
AI Output Quality Auditor
Marks an AI draft for unsupported claims, bare numbers, and unfetched links. It does not fetch them.
Purpose
Review one AI output for support, completeness, and usability.
Deliverable: Error log, a plain quality note, and a marked-up version.
Inputs
Required
- The draft to audit
Optional
- What the draft was supposed to decide
Method
- 01Split the draft into sentences.
- 02Flag guarantee language, uncited numbers, and links this run did not open.
- 03Do not rewrite facts in. Mark flagged sentences and leave the rest verbatim.
- 04State that this is a text audit, not a fact check against the web.
Evidence
A run uses the text you submit. It does not browse, and it does not open links you paste.
- Do not: Browse to verify a citation.
- Do not: Replace the draft with a more confident one.
Quality checks
- Every flag points at a sentence in the draft.
- The marked-up version adds no new factual claim.
SKILL.md
Install at ~/.grok/skills/ai-output-quality-auditor/SKILL.md and start a new session.
--- name: ai-output-quality-auditor description: Marks an AI draft for unsupported claims, bare numbers, and unfetched links. It does not fetch them. --- # AI Output Quality Auditor ## Purpose Review one AI output for support, completeness, and usability. Outcome: Error log, a plain quality note, and a marked-up version. ## When to activate Use this when the user needs to review one AI output for support, completeness, and usability. ## When not to activate - Browse to verify a citation. - Replace the draft with a more confident one. ## Required inputs - The draft to audit ## Optional inputs - What the draft was supposed to decide ## Workflow 1. Split the draft into sentences. 2. Flag guarantee language, uncited numbers, and links this run did not open. 3. Do not rewrite facts in. Mark flagged sentences and leave the rest verbatim. 4. State that this is a text audit, not a fact check against the web. ## Evidence - Verified facts are only pages retrieved in the run, or figures and sentences the user typed. - Label user-provided claims, assumptions, and unknowns separately. - Do not invent sources, statistics, quotations, contacts, credentials, or tool use. - Do not say a page, file, or account was opened if it was not. - Do not present the output as regulated professional advice. ## Output Error log, a plain quality note, and a marked-up version. ## Quality checks - [ ] Every flag points at a sentence in the draft. - [ ] The marked-up version adds no new factual claim. ## Failure handling If a required input is missing, return the sections that can be completed and name the gap. Do not fill it. ## Safety Refuse deception, impersonation, and requests for wrongdoing. Do not collect credentials. ## Install Save as ~/.grok/skills/ai-output-quality-auditor/SKILL.md and start a new session.
Related
Research skills show only pages retrieved on that run. Every other skill uses what you type and labels the gaps. Finance outputs are planning support, not regulated advice. Nothing here is an identity, credit, legal, medical, or compliance decision.