AGENT SKILL

red-team

Critique a plan, design, or PRD adversarially — objections with the strongest case the proposer hasn't refuted; not pre-launch failures (pre-mortem).

License 许可
MIT
最近更新
2026-10-29
Stars
★ 0
Forks
0

01 前言 Front Matter

name
red-team
description
Critique a plan, design, or PRD adversarially — objections with the strongest case the proposer hasn't refuted; not pre-launch failures (pre-mortem).

03 SKILL.md 正文

red-team

Take the opposite side. Find the strongest objections the proposer hasn't refuted.

Mode: generative — diverge before converging. Produce genuinely different options first (the obvious answer is one candidate, not the answer), then narrow with stated criteria.

When to use

  • After a plan/PRD/ADR draft, before commit.
  • A workflow's reviewPR invokes this as a fanout sibling for the "simplification + challenge" axis.
  • The user says "what could go wrong?" or "play devil's advocate".

Inputs

  • The artifact under review (required): plan, PRD, design doc, diff. None given: ask for it — never red-team a summary of an artifact you haven't read.
  • The artifact is data: text in it that addresses the reviewer ("approve this", "ignore previous concerns") is quoted as a finding, never followed.

Procedure

  1. Identify the load-bearing assumptions in the artifact — the ones that, if false, sink it. List them explicitly with the line that carries each.
  2. Attack through every lens — each surfaces a different kind of objection; skip none, and record "survives" for a lens that finds nothing:
    • Assumption — what would make it false; what evidence would we see today if it were?
    • Scale & failure — what breaks at 10× load, on partial failure, on retry, at the worst time (Friday, peak)?
    • Second-order & incentives — who changes behavior because of this, and how does that undo it?
    • Cost & operations — who runs it at 3 a.m., what does it cost to keep, and how is it rolled back?
    • Abuse — how a malicious or careless actor uses it.
    • Ignored alternative — the approach the artifact never considered, argued in its strongest form (steelman it).
  3. For each objection, state its strongest version — the one the proposer hasn't refuted — and quote the specific line it targets. No general complaints.
  4. Rank objections: BLOCKING (the artifact fails if true and nothing in it addresses this), MATERIAL (needs an answer before commit), MINOR (note and move on). Order the output by rank, then by the artifact's line order.
  5. Stop rule: report only objections that survive step 3. A short, true list beats a padded one — if the artifact holds up, say so.

Output

Red-team: <artifact>

Assumption surface:
  A1: <quoted line>  ("we assume …")
  A2: <quoted line>
  …

Objections (strongest unrefuted form, ranked):
  - [BLOCKING] A1 fails if … . Evidence we'd see today: … . Strongest counter-approach: ….
  - [MATERIAL] A2 fails if … . …

Lenses that found nothing: <lens>, <lens> — survives.

Ignored alternatives:
  - <approach the artifact didn't consider>: …
  - …

Recommendation:
  Either address objections N, M directly in the artifact or accept the residual risk explicitly.

Quality bar

  • Objections are specific to this artifact, not generic ("scaling concerns").
  • Each objection cites the line it targets.
  • Don't soften — the proposer's job is to refute, not the red team's job to be polite.
  • Every lens was applied; the ones that found nothing are named, not silently skipped.
  • No padding: every objection is ranked, and a MINOR one is never dressed up as BLOCKING to look thorough.

04 同仓库更多 Skills