Stage-bundles P1-4 of 4. Design: docs/design/stage-bundles.md (PR #144) §5, S4. Tracked under #137, depends on P1-3.
Scope
- Clicking a ghost row or the cell header's "how to ask" → generate 2–3 suggested questions on the spot (P1 is the out-of-meeting form: "how to open the next meeting"; the input does not include the transcript tail)
- Input: the slot hint + the claims already in that slot (so we don't ask about what we already know) + how the attendees are addressed (the buying committee)
- Output reuses the
SolutionReply shape (reply = a zh-TW business question you can read aloud verbatim; consideration = a one-liner on what the question is fishing for; kind reuses the wargame categories, or add a probe kind)
- Hard rules on phrasing, written into the prompt: build on what's already known, prefer open-ended, one question at a time, don't sound like a survey. A canned question ("So what's your budget?") is a failure case — include good and bad examples in the prompt
- UI reuses the FindingSolution card styling and lazy cache (per slot × hash of that slot's claims)
Acceptance
- Test against the AI3 / Sheraton claim libraries: would you actually dare read the generated question out loud? (Brandon's subjective sign-off, design §9 metric)
- Reopening the same slot doesn't re-hit the AI (cache); degrade gracefully when no provider key is configured (hide the entry point)
Depends on P1-3. The live form (picking up from the transcript tail) is P2.
Stage-bundles P1-4 of 4. Design:
docs/design/stage-bundles.md(PR #144) §5, S4. Tracked under #137, depends on P1-3.Scope
SolutionReplyshape (reply = a zh-TW business question you can read aloud verbatim; consideration = a one-liner on what the question is fishing for; kind reuses the wargame categories, or add aprobekind)Acceptance
Depends on P1-3. The live form (picking up from the transcript tail) is P2.