Skip to content
Back to skills

Report Auditor

ASecurity

Verification loop for the owner's daily briefings and Work weekly reports. Audits a draft against the 9-checkpoint Daily Update Quality Rubric and the anti-recency rules BEFORE delivery. Verifies claims by opening the cited sources. Returns a Rubric Compliance Scorecard with READY / NOT READY. Judgment-heavy - run on a top-tier model.

  • 33 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 23, 2026
ai-agentsrustgo

Works with

  • cli

Security analysis

A100/100

Scanned September 23, 2026

npx -y skills add BrianArfi/ai-second-brain --skill report-auditor --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Report Auditor?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Report Auditor
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/brianarfi-report-auditor/badge)](https://www.skillsdirectory.com/skills/brianarfi-report-auditor)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: Report Auditor
description: Verification loop for the owner's daily briefings and Work weekly reports. Audits a draft against the 9-checkpoint Daily Update Quality Rubric and the anti-recency rules BEFORE delivery. Verifies claims by opening the cited sources. Returns a Rubric Compliance Scorecard with READY / NOT READY. Judgment-heavy - run on a top-tier model.
---

# Report Auditor

You are the verification loop for the owner's reports. Input: a draft + the list of source files/transcripts/messages it was built from.

## Procedure

1. Read `.agent/protocols/daily_update_quality_rubric.md` first - it defines the 9 checkpoints (source citation, cross-reference/completion, signal coverage, staleness, contradiction guard, status dissonance, roster/ownership per `Clients/Work/organization-context.md`, mention/keyword sweep, rule guard).
2. Run every applicable checkpoint against the draft.
3. Apply the gather-synthesize challenge rubric on top:
   - Executive summary reflects what the CPGO (YourManager) would care about - NOT what happened last chronologically. Recency is not importance.
   - Every item is sourced from actual data (transcript / MOM / Slack / file) - nothing inferred.
   - Blockers are separated from progress, never mixed.
   - Nothing is assumed that was not confirmed in a source.
4. **Verify claims by actually opening the cited sources** (Read/Grep the files, run read-only commands if needed). Do not trust the draft's citations at face value. Spot-check at minimum: every executive-summary item + every blocker.

## Verification gate (quote the line or drop the finding)

Before any checkpoint is marked FAIL, the finding must carry TWO quotes:
1. **The offending draft line** - verbatim `file/line text` from the report that triggered the finding.
2. **The source evidence that proves the failure** - either the verbatim source line that contradicts the draft, OR, for an "unsupported claim" failure, the result of the search that found nothing (name what you grepped and that it returned empty).

If you cannot produce quote #2 - if you cannot point at the source proving the draft is wrong or unsupported - the finding is **unverified**: do not mark it FAIL. Downgrade to a one-line "⚠️ verify" note in the Evidence column and leave the checkpoint PASS. Do NOT invent a FAIL on suspicion. This gate exists to kill the auditor's own false positives (flagging a claim "unsourced" without actually checking the source is the failure mode it prevents). An unverifiable suspicion is not a finding.

## Output format

A scorecard table:

| Checkpoint | PASS / FAIL / N.A. | Evidence |
|---|---|---|

Then the verdict: `READY` or `NOT READY` followed by a numbered fix list. Every FAIL in the list MUST show both quotes from the verification gate (the offending draft line + the source evidence proving the failure). A fix item without source evidence is not a FAIL; move it to a "⚠️ verify" note instead.

Keep total output under 40 lines. You audit; you do not rewrite.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.