| name | run-challenger |
| description | Use before requesting G2 approval for an EXP plan, or when scope changes are proposed mid-Production (Rule 11 / Challenger role). Trigger phrases: "run challenger", "challenge this plan", "stress-test the approach", "draft the challenger block", "what could go wrong with this plan". Produces the four mandatory Challenger sections per [02-exploration/challenger.md]. |
Run Challenger
The Challenger is the AI's internal sceptic. It moves friction into the AI — where it is cheap — instead of relying on the human's vigilance at the gate. The Challenger does not block; it makes the cost of being wrong visible before G2 is crossed.
When to trigger
- Mandatory — Before AI requests G2 approval; plan is transitioning
draft → proposed and has no ## Challenger section yet
- Conditional — During Production when the human proposes a scope expansion
- Optional — During Context when
risks.md is suspiciously thin (< 3 risks or all rated low)
Prerequisite
The target EXP plan exists and has, at minimum, a stated approach, alternatives, and a ## Vision Alignment section (or an explicit — with a reason). If those are missing, redirect to create-exp-plan first.
Steps
- Read the role spec: 02-exploration/challenger.md. The four mandatory sections and the auto-fail anti-patterns are defined there.
- Read source material: the target EXP plan, its
hypotheses.md, risks.md, related ADRs, and any prior approved plans this proposal builds on (for decision continuity).
- Draft the four sections — each must be specific to this plan, not generic:
- Top-3 failure modes — for each: what breaks (a concrete observable, not "things go wrong"), the cheapest signal that would tell us early, what we would do
- Two alternative approaches — sketch (3–5 lines), why rejected, the condition under which we would re-open
- Strongest counter-argument — first-person, charitable, the sharpest reviewer's objection
- Product-coherence check — vision fit, decision continuity, problem validation (measured vs. hypothetical)
- Self-check against the auto-fail patterns before showing the user:
- Generic failure modes ("schedule slip", "scope creep") → regenerate
- Alternatives that are "do nothing" or "use a different framework" → regenerate
- Strawman counter-argument → regenerate
- "Fine, aligned, measured" coherence paragraph with no specifics → regenerate
- Plan has no
## Vision Alignment section → record this as a finding for G2; do not paper over it
- Place the output in the plan file under
## Challenger, or for very large plans in a sibling file <plan-id>.challenger.md.
- Confirm the draft with the user before writing.
Rules
- The Challenger does not issue verdicts — that is the Gatekeeper agents' job. The Challenger surfaces the cost of being wrong; the human decides.
- Code review is not the Challenger's job — that happens at G3 against acceptance criteria.
- The Challenger is not the human's voice — humans review Challenger output, they do not write it.
- If a prior Challenger failure mode came true during Production, append it to
insights.md (write-insights skill) — the loop closes.
Output
A ## Challenger block (in the plan, or a sibling file for very large plans) containing all four sections, each non-generic and concretely tied to the plan. The Gatekeeper checks both existence and substance before issuing G2 pass.