cowork-review

name	cowork-review
description	Orchestrate a full multi-agent document review pipeline in one pass — load, evaluate, verify, draft, and audit.
argument-hint	[archive-path]
disable-model-invocation	true

Document Review Pipeline

Orchestrate a full multi-agent document review: load, evaluate, verify, draft, and audit — in one pass.

Instructions

1. Gather all evidence (one command)

Run the pipeline script to load documents, evaluate against all criteria files, and collect the audit trail:

uv run --with 'pymupdf,python-docx,openpyxl,chromadb,sentence-transformers' \
    "${CLAUDE_PLUGIN_ROOT}/skills/cowork-review/document-pipeline.py" "$ARGUMENTS"

Default archive: assets/ Default criteria: ip-and-ownership.md+general-red-flags.md`

If data is already loaded, add --skip-load to skip the load stage.

2. Assess findings

Read the evidence output. For each criterion, apply judgment:

• Rate severity: CRITICAL, HIGH, MEDIUM, LOW, or CLEAR
• Quote specific clause text as evidence
• Note source document(s) and section numbers

3. Adversarial review

For each finding rated MEDIUM or above, switch to verification-agent perspective:

• Construct the best counter-argument
• Search for exculpatory context
• Verdict: upheld, downgraded, or dismissed

4. Draft response

Switch to response-drafter-agent perspective:

• Group verified findings by severity and topic
• Draft a professional memo: executive summary, critical issues, high-priority concerns, moderate items, positive notes, next steps
• Tone: collaborative, specific, constructive

5. Save and display

Save the full report to data/review-report.md. Display the audit trail summary at the end.

Options

• --skip-load — skip load stage if data already exists
• --criteria FILE ... — use custom criteria files instead of defaults

Constraints

• The pipeline script handles load + eval + audit in one permission grant
• The LLM handles assessment + adversarial + drafting (judgment work)
• Always cite section numbers, source documents, and quote evidence
• Save final report to data/review-report.md