Grill Voice

Grill by Voice

Run the mp-grill interview, but exchange each round with the user's mobile voice app through JSON files instead of conversation. The user is on a walk: questions are spoken to them, answers come back as Whisper transcripts. The file formats and session lifecycle live in CONTRACT.md — read it before the first round.

All session bookkeeping goes through one script:

node "${CLAUDE_SKILL_DIR}/scripts/grill-voice.js" <init|publish|wait|complete> ...

Step 1: Context and subject

Follow mp-grill Steps 1–2: silently read .mpx/CONTEXT.md and .mpx/DECISIONS.md when present, and resolve $ARGUMENTS as the grilling subject (ask when absent).

Step 2: Start the session

node "${CLAUDE_SKILL_DIR}/scripts/grill-voice.js" init --project <name> --topic <topic>

Prints the sessionId used by every later command. Confirm to the user that the session is live and they can put the phone in their pocket.

Fallback to conversational grilling — when the user says they are at the keyboard, or the script fails (no Node, unwritable sessions root): continue with mp-grill Step 3 in the conversation and skip the publish/wait cycle entirely.

Step 3: Rounds

Compose each round exactly as mp-grill Step 3 prescribes: delegate codebase facts to the Explore agent (breadth: medium), batch related questions thematically, split into a follow-up round only when earlier answers materially change later questions, and attach a recommendation to every question.

Voice changes only the delivery:

1. Write questions for the ear — short sentences, no markdown or code syntax, identifiers spelled out. Include the round announcement naming project, topic, and round number so the user can tell sessions apart (formats in CONTRACT.md).

2. Write the round JSON to the session scratchpad, then publish it:

   node "${CLAUDE_SKILL_DIR}/scripts/grill-voice.js" publish <sessionId> <roundFile>
   

3. Wait for the spoken answers:

   node "${CLAUDE_SKILL_DIR}/scripts/grill-voice.js" wait <sessionId> <round>
   

   Exit 0 prints the answers JSON. Exit 2 means still waiting — run wait again, indefinitely; the user answers at walking pace, and each wait call polls for several minutes before returning.

4. Treat each transcript as the user's answer. Transcripts are speech: read them charitably (homophones, spelled-out identifiers) and carry any genuinely ambiguous transcript into the next round as a clarification question. A skipped question is re-asked once in a later round or explicitly dropped with a note in the report.

Step 4: Conclude

Follow mp-grill Step 4 for CONTEXT.md / DECISIONS.md updates, with one adaptation: where mp-grill would ask the user whether an uncertain entry belongs in the docs, put those confirmations into one final voice round instead of asking in conversation. Each confirmation question speaks the full candidate entry — for a Domain Language term, the term and its complete one-sentence definition, voice-adapted per CONTRACT.md — so the user hears exactly what would be written before it lands in the docs.

Then close the session so it leaves the app's active list:

node "${CLAUDE_SKILL_DIR}/scripts/grill-voice.js" complete <sessionId>

Gate: Continue only when the completion command succeeds and the session is absent from the app's active list.

Report

Summarize as mp-grill does — decisions made, requirements clarified, docs updated, open items — and note anything lost to voice: skipped questions, ambiguous transcripts, and where each was resolved or dropped.