/challenge - Red-Team Review

Runtime	~5 minutes
Reads	Target document, theme claude.md, people.md, vault history
Writes	Challenge report to stdout, fixes to tasks.md
Model	Opus (5 parallel subagents)

What It Does

Stress-tests any document, plan, or idea through five parallel analysis lenses before it leaves the building. Returns a verdict, the top issues to fix, and a pre-mortem of how it fails.

Why It Matters

Every important document has blind spots. The author can't see them because they built the argument. A single reviewer applies one perspective. Serial critique is slow and tends to anchor on the first problem found.

/challenge runs five independent reviewers simultaneously, each looking through a different lens. The result arrives in one pass, not five rounds of "what do you think?"

How It Works

graph TD
    A[Target Document] --> B[Context Gathering]
    B --> C1[Audience Fit]
    B --> C2[Logic & Evidence]
    B --> C3[Vault Contradictions]
    B --> C4[Known Blind Spots]
    B --> C5[Pre-Mortem]
    C1 --> D[Consolidated Challenge Report]
    C2 --> D
    C3 --> D
    C4 --> D
    C5 --> D
    D --> E[Top 3 Issues + Verdict]

1. Identify target - Accepts a file path, topic keyword, or (with no arguments) challenges whatever you've been working on this session
2. Gather context - Reads theme files, stakeholder profiles, and previous versions to understand what the document is trying to do and for whom
3. Launch five parallel subagents - Each receives the full document plus context, each applies its specific lens
4. Consolidate - Merges findings into a single report with a verdict (Strong / Needs Work / Weak)
5. Extract tasks - Any actionable fixes go straight into tasks.md

The Key Innovation

Parallel lenses, not sequential critique. The five subagents run simultaneously as Task workers:

1. Audience Fit - Would the target reader actually finish this? Does it match their stated communication preferences from people.md? Is the ask clear in the first 30 seconds?
2. Logic & Evidence - Are claims backed by specific numbers or just directional language? Would a sceptic find it compelling?
3. Vault Contradictions - Searches the vault for conflicts. Has the positioning shifted without acknowledging the change? Do prior meeting notes contradict what's being proposed?
4. Known Blind Spots - Applies your documented blind spots from CLAUDE.md. Architecture without execution mechanics? Culture angle missing? Greenfield bias showing?
5. Pre-Mortem - How does this fail? Top 3 failure modes, strongest counter-argument, the question your audience will ask that isn't answered.

This matters because a single sequential review anchors on one perspective. Parallel execution means the pre-mortem doesn't know what the audience-fit lens found, and vice versa. Independent analysis produces genuinely different angles.

Example Usage

/challenge 02_Themes/project-a/emails/2026-02-25_board-update.md

Or challenge by topic:

/challenge board presentation on AI transformation ROI

Or challenge what you've been working on:

/challenge

Output includes a verdict, the top 3 fixes, and specific sections for each lens. The "What's Actually Good" section prevents the review from feeling purely negative - but it never contains filler praise.

Customisation Guide

• Blind spots - The fourth lens pulls from your Known Blind Spots section in CLAUDE.md. The more specific your blind spots, the more useful this lens becomes. Review quarterly.
• Stakeholder profiles - Audience Fit checks people.md for communication preferences. Richer profiles produce sharper audience analysis.
• Vault depth - The Vault Contradictions lens searches meeting notes, emails, and status docs. The more history in your vault, the more contradictions it catches.
• Stakes calibration - The skill scales effort to consequence. A Slack message gets a lighter treatment than a board paper. You can override by saying "full challenge" or "quick challenge".