grilling

Forks: read Poteto.

Resolve unknowns; no production code or implementation while a material user-reserved decision is open. Invocation grants no delegation.

Plans/RFCs/replies: read communication. Load skills per dependency loading.

1. Build evidence

Ground request, plan and decisions in repo/tests/docs. Ask only for user-owned preferences, scope, risk appetite or decisions evidence cannot settle. Name the blind spot. Prototype when behavior beats prose.

/brain-dump is optional: retain every opportunity track and its Answer ledger. Never re-ask Settled; challenge Tentative only for material downside; ask Unknown only if it could invalidate/prioritize a track.

2. Explore mode

Interface/ownership fork: load "codebase-design"; sketch caller use.

Without direction, compare 2-3 approaches with trade-offs, reversibility, evidence and a recommendation. For competing plans, Call the Skill tool with "plan-arbiter". Challenge variant: with direction, steelman the best alternative and its disproof.

For customer-facing UI, show an ASCII wireframe before questions: each materially different layout in fenced text, with real labels, controls, grouping, order, and fixed/scroll regions. Align borders; show structure, not pixels. Desktop/mobile only when composition differs; shared layouts use one sketch plus deltas.

Map the decision tree's answerable frontier. Ask the whole frontier in one numbered round using this Question format:

**Q1 -- <question title>**
<question or choices>

**Recommended:** <answer>

---

**Q2 -- <question title>**
<question or choices>

A confident, reversible Recommended: answer is an assumption, not a question: state it in one line and continue. Ask only what evidence cannot settle and the user owns.

Stop signal: "no more questions", "just do it", "don't ask", or a bare endpoint ("draft PR") ends asking for the rest of the session. Adopt every open recommendation, list them as assumptions, and continue to the endpoint; pause only for an irreversible or user-reserved decision.

Ask so yes accepts the recommendation.

An unsettled prerequisite delays only its branch. Recompute the frontier each round. Keep fact-finding inline; explicit delegation uses fresh contexts and self-contained briefs. Use environment/filesystem/tools. User decisions stay theirs.

3. Exit

Resolve or reserve architecture-changing choices. Classify the rest as lookup -> prototype -> reversible assumption -> pause trigger. Exit when nothing unresolved can silently invalidate the next slice.

4. Plan gate

Build one Evidence packet: request, plan, spec/standards sources, paths, repo facts, assumptions, unresolved decisions.

• Quick: under three tasks, no material architecture/product/UX choice; check spec, standards, value inline.
• Standard: product/spec, engineering/standards, and design/UX reviewer hats inline.
• Deep-risk: Standard plus separate Skill tool calls for "resilience-review" and "steelman" for a credible high-impact or hard-to-reverse assumption.

Axes: Spec -> plan-product-hat; Standards -> plan-engineering-hat; design/UX -> plan-design-hat; plus adversarial/value. Deep-risk triggers: auth, migration, public API, destructive action, concurrency, Temporal, cross-service work, and one-way doors.

Specialist registry: Call the Skill tool separately with "product" for value and "golang" for planned Go or go.mod work; add specialists after repeated misses. Axes report APPROVED, NEEDS_CHANGES, BLOCKED, or SKIPPED with evidence/skip reason. Dedupe root causes; research facts; stop on blocking user decisions.

Confirm only for a requested plan/grill endpoint. [ETHOS: Discover Before Commitment]

Call the Skill tool with "domain-modeling" for GLOSSARY.md terms; ADR only for hard-to-reverse, surprising trade-offs.