Strategy Red-Team (attack the assumptions with counter-evidence)

Be a sharp, fair adversary against a strategy/PRD/roadmap — but a grounded one. The differentiator over a generic red-team: every attack points to real customer counter-evidence where it exists. Not "this might be risky" but "the doc assumes X; three customers said the opposite 1." Your grounding searches are designed to falsify the document's claims, not confirm them.

Step 0 — Relevance & availability

Confirm there's a concrete document to attack and Evermuse is connected. If not connected, run a framework-only red-team labeled ⚠ ungrounded — Evermuse not connected (assumption logic only, no counter-evidence) and tell the user to authorize the MCP. See ${CLAUDE_PLUGIN_ROOT}/skills/using-evermuse/SKILL.md Step 0.

Mode

• Red-team (default): attack load-bearing assumptions now, while the cheapest test is still available.
• Pre-mortem mode: if the user asks to "run a pre-mortem" or imagine the launch already failed, use references/pre-mortem-template.md (Tigers / Paper Tigers / Elephants → launch-blocking / fast-follow / track). Same counter-evidence grounding applies.

Evermuse Grounding (required)

Follow ${CLAUDE_PLUGIN_ROOT}/skills/using-evermuse/SKILL.md.

Required — one parallel batch of searches. Verify the product (Rule 1), then fire in a single parallel batch:

• 3–4 evidence searches, each worded from a different angle (limit up to 50). Evidence comes back rich and varied — expect large, useful result sets.
• one guidance search and one context search (limit up to 50). These are usually sparse or empty; run them anyway and note when they're thin.

Read each response's digest — it reports how many more results exist. Use judgment on whether a query is worth pulling deeper (raise limit toward the 100 max and/or page with next_offset to avoid repeats), weighing payload size, remaining context, task complexity, and the value of the data. For deep pulls, consider spawning sub-agents — instruct them to return every citation with the same metadata the tools return (url, who_said_it, meeting_name, created_at) so you can still cite.

Optional — considered use. Once grounded, reach for the other tools only when they add value: a quote-angled search for verbatim voice, read_source for a single deep dive, inline citations (references/citations.md), and add_source to save the deliverable. Sources, citations, and saving are optional — not required.

Instructions

You are a sharp, fair adversary reviewing $ARGUMENTS.

1. Extract every claim. List what the doc asserts as true. Separate load-bearing (if false, the plan dies) from cosmetic. Only load-bearing claims are worth attacking.
2. Design falsifying searches. For each load-bearing claim, phrase an Evermuse search that would surface the opposite — the customers who'd disagree. This is the core move.
3. Steelman, then attack. State the strongest version of the claim; then attack that, anchored in the counter-evidence found. No strawmen.
4. Write each failure mode as "Fails if ___." Concrete and falsifiable. "Fails if activation isn't actually the constraint" beats "execution risk."
5. Rank by (impact if wrong) × (likelihood wrong) × (cheapness to test). Surface the top of the list — the highest-impact, plausibly-wrong, cheap-to-check assumption is what to test this week.
6. For each surviving kill-assumption, give the operator something to do: Fails if · Counter-evidence (the opposing quotes, cited) · Evidence to get this week · Kill criterion (the threshold to stop/change) · Cheapest test.
7. Self-refute, don't fabricate. Default to "this risk is real" unless a falsifying search comes back empty or the doc already cites evidence against it — then say the claim holds.

Where the source method reaches for "web search / market research," use Evermuse evidence (falsifying customer voice) as the primary source and context (market) as secondary.

Deliverable format

## Red-Team: [plan in one line]

### Top Kill-Assumptions (ranked)
For each (3–5 max):
- **Claim:** [the load-bearing assertion in the doc]
- **Counter-evidence:** > "[opposing quote]" — [attribution] [`1`](URL)  (you assume X; [N] customers said the opposite)
- **Fails if:** [concrete, falsifiable condition]
- **Evidence to get this week:** [specific]
- **Kill criterion:** [threshold]
- **Cheapest test:** [smallest experiment]

### What's Well-Reasoned
[Claims that survived a falsifying search — state why, cite the confirming evidence.]

### What I Couldn't Assess
[Gaps where neither the doc nor Evermuse gave enough to judge.]

End with what to do, not just what to fear — the emotional job is relief from shipping the wrong bet.

Further reading

• Adapted from phuryn/pm-skills (MIT), merging its strategy-red-team and pre-mortem skills; grounding is Evermuse-native.
• Assumption Prioritization Canvas
• How Meta and Instagram Use Pre-Mortems