grill-me

Grill Me

Interview the user until consequential design choices are explicit, then write a Design Concept that downstream SpecKit skills can use without reinterpreting the conversation.

Ground recommendations

Inspect the tools and skills actually available. Follow the shared capability-discovery and grounding contracts. Base recommendations on the project constitution, codebase evidence, or current authoritative sources; disclose uncertainty instead of guessing.

Interactive boundary

Allowed entry points are an active user requesting Grill Me by natural language or invoking $grill-me, and an interactive $speckit-scaffold-spec call. Before any question or write, confirm this is an active user chat that can receive a direct reply.

Abort in background or non-interactive execution, codex exec, CI, autopilot, any phase or consensus agent, and every subagent context. Say that Grill Me requires an active user conversation and that autopilot uses its Clarify consensus flow. Do not ask a question and do not write any file.

Codex interaction adapter

Prefer request_user_input whenever it is present. Send exactly one question, 2-3 mutually exclusive choices, and the grounded recommendation first with the label suffix (Recommended). Give each choice a concise tradeoff and wait for the user's reply before continuing.

If the picker is absent or its call is unavailable, a free-text fallback is allowed only in the already active user chat. Ask exactly one question in the current conversation, list the recommended choice first plus 1-2 mutually exclusive alternatives with tradeoffs, and wait for the user's direct reply. Never use this fallback in background, CI, autopilot, or subagent execution.

Workflow

1. Determine the mode and input:
   • Standalone: accept a file, topic, or interactive input; propose docs/ai/specs/<slug>-design-concept.md unless the user supplied a path.
   • Setup: use the scope and output path supplied by $speckit-scaffold-spec; never redirect the write to the primary checkout.
2. Read the shared interview protocol. Ground the initial model in applicable project instructions, constitution, roadmap, prior design decisions, and targeted code. If docs/ai/specs/ubiquitous-language.md exists, read it before the Terms branch and reuse its rows.
3. Walk the highest-impact, highest-uncertainty design branch first. Ask one neutral decision question, record the recommendation and evidence, record the user's answer, and update the remaining branches.
4. Include a slice-sizing branch near the end. Read the canonical slicing heuristics, derive story, surface, requirement, and new-versus-modify signals, and run runner operation estimate-spec-size.
   • Treat warn or a horizontal slice as a reason to recommend thin vertical slices, never as a gate.
   • Treat an unavailable, non-zero, empty, or unparseable estimate as absent; note it and continue.
   • Record an accepted split in Goals, a deferred split in Open Questions, and a declined or unnecessary split as an advisory note.
5. Stop at natural convergence, when the user ends the interview, or at the protocol's cap. Only after the interview, read the shared Design Concept output contract and synthesize the record.