grilling
Stress-test a product plan, decision, or idea through a relentless one-question-at-a-time interview. Use when the user asks to be grilled, wants assumptions challenged, has a vague or risky plan that
SKILL.mdplugins/maga/skills/grilling/SKILL.md
GitHubGrilling
Map the uncertainty as a design tree: each product decision branches into the decisions that depend on it. The frontier contains decisions whose prerequisites are already settled. Recompute it after every answer so hidden assumptions become visible without asking downstream questions too early.
Investigate facts from the environment or authoritative sources instead of asking the user. Independent fact lookups may run in parallel while decisions that do not depend on them continue. Product decisions remain the user's.
Ask exactly one highest-leverage frontier question at a time and recommend an answer with its product consequence. This is MAGA's deliberate adaptation of upstream round-based grilling: product builders should not receive an entire frontier as a questionnaire. Only when the user explicitly asks for a batch and the decisions are independent may one turn contain up to three numbered questions.
Choose The Mode Automatically
- Standalone: use for an idea or plan with no repository, or when the user only wants conversational challenge. Do not create project documents.
- Documented: use when the work belongs to a repository and resolved terms
or hard-to-reverse decisions should survive the conversation. Apply
domain-modelingas needed to update the repository's established context or decision records; do not invent a parallel document system.
Do not ask the user to choose a mode or invoke another Skill. Switch from standalone to documented only when durable project context becomes useful.
Finish Deliberately
Do not implement the plan during the interview. Stop when the current useful product slice has no unresolved frontier decision—not when every hypothetical future branch has been explored. Summarize the shared understanding, remaining uncertainty, and recommended next product step. Continue into action only after the user has authorized that next step or their original request already did so.