Strategy Red Team · free MIT Skill
Attack the plan before reality does.
Hand your assistant a PRD, roadmap or strategy. It finds the claims the plan cannot survive without, attacks the strongest version of each, and tells you the cheapest test to run this week and the threshold at which to stop.
Right: an unedited excerpt from a saved run on a fictional lunch-kit startup.
Claim: Seattle will behave like Portland (and justifies a second kitchen in Q2).
Steelman: the product is built for a segment (20–45 person studios without kitchens) that exists in any tech/design city, and Seattle has more of them than Portland.
Fails if: Seattle demand, pricing and occupancy differ enough that a second kitchen sits underused… The plan supplies no Seattle data at all.
Cheapest test: Seattle waitlist plus pre-order via the existing Portland kitchen on a small trial route, rather than capital for a second kitchen.
Fewer than ~10 Seattle offices (est. threshold) commit to a trial or pre-order at $11.50 before any lease is signed.
cosmetic, not attacked: ad creative, add-on naming, exact Wednesday cutoff
What it is
A single instruction file that turns a capable AI assistant into a sharp, fair adversary. It lists everything the plan asserts, keeps only the load-bearing claims, steelmans each one and then attacks that strongest version — never a strawman.
Every surviving risk is written as “Fails if…”, ranked by impact × likelihood × cheapness to test, and paired with evidence to get this week, a kill criterion and the cheapest test. If part of the plan is sound, the Skill must say so.
Who it is for
- Product managers taking a PRD or roadmap into executive review.
- Founders about to commit capital, a hire or a launch date.
- Strategy leads who want a challenger that is specific, not generic.
- Reviewers asked to “poke holes” in someone else’s plan, fairly.
Send it
- A strategy, PRD or roadmap before a go / no-go meeting
- Expansion, pricing, launch or new-feature bets
- A plan that has only had polite feedback so far
- A second opinion from another model, if you ask for it
Send it elsewhere
- Narrative “how did this fail” workshops — use a pre-mortem
- Exhaustive risk registers; it keeps 3–5 kill-assumptions
- Financial, legal or compliance sign-off
- Fresh market data; it does not browse on its own
What you provide, what you receive
You provide
- The plan: PRD, roadmap, strategy memo or a few lines
- The numbers it rests on (traction, margin, goals)
- Any evidence you already have for its claims
- Optional: ask for a cross-model second opinion
You receive
- Red-Team: the plan in one line
- 3–5 ranked kill-assumptions, each with Claim · Fails if · Evidence to get this week · Kill criterion · Cheapest test
- What’s Well-Reasoned — what holds up, and why
- What I Couldn’t Assess — gaps in the plan
Where it runs
Any AI assistant that can read a Skill file, including Claude Code and Codex. Download the free plaintext package from Finch, or the source from GitHub, then install it with your assistant’s Skill method or ask it to read SKILL.md.
No extra API key, script or service. You still need access to an assistant model, which may carry its own subscription or usage cost. The Skill has no built-in web research; it judges the plan and evidence you give it. The optional cross-model mode needs a second model you can already reach.
Limits and licence
Output quality depends on the plan and the model. Thresholds it proposes are judgments to confirm, not measured facts.
MIT licence, Copyright (c) 2026 Pawel Huryn. Source: phuryn/pm-skills at commit 2a12295a7717, path pm-execution/skills/strategy-red-team. The package adds a README, NOTICE, provenance note and a promotional cover; the instructions are unchanged.