Skip to content

claude-md-snippet.md

Latest commit

 

History

History
77 lines (62 loc) · 3.01 KB

claude-md-snippet.md

File metadata and controls

77 lines (62 loc) · 3.01 KB

Paste this into your project's CLAUDE.md

The block below is the entire setup. No install, no plugin, no marketplace, no npm, no pip. Open your project's CLAUDE.md (the file Claude Code reads on every session) and append the section between the --- SNIP --- lines.

You'll need Claude Code, Codex CLI, and Gemini CLI on your PATH already — that's the only prerequisite. Chapter 0 walks you through those if you don't have them yet.

--- SNIP --- (start) ---

## Multi-model design red team

When reviewing a plan, spec, RFC, or design document — especially
one with distributed state, IAM, data integrity, or significant
blast radius — run a multi-LLM red team review.

1. Use `run_in_background` to dispatch the same prompt + plan to
   Codex CLI and Gemini CLI in parallel. Concretely, write the
   `<prompt + plan>` text to a temp file and run:
   - `codex exec --skip-git-repo-check < /tmp/redteam-input.txt > /tmp/codex.md 2>&1`
   - `gemini < /tmp/redteam-input.txt > /tmp/gemini.md 2>&1`
2. While they run, do your own review with the same prompt and write
   to `/tmp/claude.md`.
3. Use the adversarial mindset only: find failure modes, do not
   propose fixes.

Prompt to give every model (yourself included):

> You are the red team for this design. For each of the 5
> dimensions below, give AT LEAST 2 concrete failure scenarios.
> Each scenario must include TRIGGER, IMPACT, and DETECTABILITY:
>
> 1. Hidden assumptions — ordering, uniqueness, atomicity, data
>    freshness, caller behavior
> 2. Dependency failures — upstream / downstream degradation
> 3. Boundary inputs — empty, huge batch, malicious, malformed
> 4. Misuse paths — caller / user / operator getting it wrong
> 5. Rollback & blast radius — how to recover, 5-minute vs 5-day
>    detection
>
> Reject abstract advice ("add monitoring"). Say what metric, what
> threshold, what alert.

After all three return, read `/tmp/{claude,codex,gemini}.md` and
consolidate into:
- Consensus (≥ 2 teams agreed)
- Unique (1 team only — could be that team's insight, or could be
  the others' blind spot)
- Apparent disagreements (escalate to user)

Severity: cap MUST-FIX at 5 unless the design is architecturally
broken. Respond in English.

--- SNIP --- (end) ---

That's it. Restart your Claude Code session if it was already running, and the next time you ask Claude to review a plan in this project, it'll fan out to all three models in parallel and bring back a consolidated report.

Why this works without an install

Claude Code reads your project's CLAUDE.md on every session. Any instruction you put there becomes part of how Claude behaves in this project. The snippet above is just instructions: "when reviewing a plan, also run these other CLIs and consolidate." No binary, no plugin API, nothing to break when Claude Code updates.

If you'd rather have a one-shot script outside Claude Code, see 06-going-further/final/redteam.sh — same method, ~100 lines of bash, takes a plan.md and writes a ranked report.