| name | grill-product |
| description | Analyze an existing product, then grill, challenge, and push back on it to make it sharper — or shape a new product idea from scratch. Use when the user wants to pressure-test, improve, or stress-test a product, decision, feature, spec, or PRD, or asks to grill a product. |
Grill Product
Make an existing product sharp enough to win, and a new product decision sharp enough to build. The goal is improvement, not interrogation for its own sake: when AI makes execution cheap, what to fix and what to build is the decision that matters.
Optimized for brownfield — analyzing and improving real, built products. It also shapes new ideas from scratch when there is no product yet.
Pick The Mode
- Existing product (primary) — there is a built product, feature, spec, PRD, repo, or live experience. Analyze it first, then grill, challenge, and push back to improve it. Do not critique before you understand what it does, who it serves, and where value is reached or lost.
- New product — only an idea or a problem. Start from scratch and grill the decision into shape.
Default to the existing-product track unless the user clearly has only an idea. When unsure, ask once.
How To Run It
- Read references/steps.md and run the track for the detected mode.
- Pull questions from references/lenses.md — only the lenses that fit the decision, not all five.
- Consult references/frameworks.md when a lens needs a named framework (MOAT, selling motion, value metric, product shape, tradeoff axes).
Core rules:
- Grill one question at a time, each with your recommended answer.
- Self-answer from the product, repo, data, or web when you can. Tag every finding verified / provided / assumed (see steps.md).
- Gather real outside voices — users, market, competitors — via web search,
gh, and other tools, fanned out to subagents (see references/outside-voices.md). Don't grill on assumptions when evidence is reachable.
- Ask the user only when a missing input would change the decision.
- Treat what users and stakeholders say as a signal to interpret, not an answer to accept.
- In a single non-interactive turn, lead with the analysis, the brief-so-far, and the top unresolved question — do not fake closure (see steps.md rhythm).
Effort Levels
Scale the grill to the ask. Default to standard.
| Level | Lenses | Outside voices | Depth | Output |
|---|
quick | 1–2 highest-risk | skip | a few questions | findings inline, no brief |
standard | 2–3 by weakness | one sweep | full grill | Sharpened Brief |
deep | all relevant | full fan-out + vet | exhaustive, iterate | Sharpened Brief + evidence |
Common Mistakes
- Grilling before understanding the product (skip analysis in Track A).
- Accepting the user's framing instead of interpreting the signal behind it.
- Running all five lenses by default instead of the ones that fit.
- Producing a long report instead of one-question-at-a-time pressure.
- Spreading pressure evenly instead of hitting the weakest part.
- Going through the motions — ending with "looks good" and no committed action.
- Faking closure instead of naming open risks and a verdict.
Scope Boundaries
Use this to analyze and improve a product decision, or to shape a new one. Route elsewhere when the request is different:
- Critique of an existing built UX surface only: use
ux-reviewer.
- Copy-only writing or review: use
ux-content or copywriting.
- Code, PR logic, or implementation review: use a code review skill.
- Generic plan or design grilling not specific to product: use
grill-me.
- Generating new UX concepts from scratch: use a concept/design skill.
