strategy-red-team acts as a sharp, fair adversary on a plan. It extracts every claim, keeps only the load-bearing ones, states the strongest version of each and then attacks that version. Each failure is written as “Fails if ___” so it can be tested.
Failure modes are ranked by impact if wrong, likelihood of being wrong and how cheap they are to test. For the three to five that survive it gives the evidence to collect this week, a kill criterion and the smallest experiment.
What You Get
- Three to five ranked kill-assumptions
- For each: the claim, the Fails if condition, the evidence to get this week, the kill criterion and the cheapest test
- A What’s Well-Reasoned section listing what survives the attack
- A What I Couldn’t Assess section for gaps in the plan
- An optional second-model comparison on request
Common Use Cases
- Pressure-testing a PRD before engineering starts
- Preparing a strategy doc for executive review
- Choosing which assumption to test first this week
What Makes It Unique
It bans strawmen, generic risk lists and invented weaknesses, and it ranks by cheapness to test so the output ends with what to do. An optional cross-model mode runs the plan through another model and flags where the two disagree.
Under The Hood
The skill is one SKILL.md of instructions with no scripts. The agent reads the plan file you provide and reasons through the claims. Only if you ask for a second opinion and another model is reachable does it send the plan to that model; the default is a single-model review that calls nothing external.