4.2 KiB
name, description
| name | description |
|---|---|
| grill | Stress-test a plan, decision, or design through a one-question-at-a-time interview. Use when the user asks to grill a plan or uses a grill trigger phrase. |
grill
Interview the user until the design tree has no open questions. Open questions define the ending condition for the grill session. The design tree tracks which decisions depend on other decisions.
Do not create glossary, context, or ADR files during grill. Do not begin implementation during or after grill without an explicit user instruction. A completed grill summary is not an instruction to build.
Run the interview
-
Investigate first. Read the conversation, files, and docs that can settle facts or reveal constraints. Resolve factual questions yourself when tools can answer them. Do not ask the user for facts you can find.
Finish when known facts, constraints, and unresolved decisions are visible.
-
Build the open-question list. Map the design tree from the known plan or idea. List the questions needed to finish the grill session. Defer questions that depend on another open question.
Finish when the open-question list defines the session ending condition.
-
Start the session. Present the full open-question list in short form. Include titles only, not full question bodies. Mark each unresolved question with
- [ ]. Then ask the first numbered question in the same turn.Finish when the user sees the open-question list and the first question.
-
Run the question loop. Show the full open-question list on every active grill turn. Ask one numbered question at a time. Apply Question format to each question. Wait for the user's answer. Apply Open-question accounting before asking again. Resolve newly answerable facts yourself before asking again.
Finish when one answer has been accounted for.
-
Close the interview. Stop when the open-question list is empty. Output a concise summary of every decision reached. Include speculative follow-up questions that were not added to the open-question list. Do not act on the summary unless the user gives a new implementation instruction.
Finish when the summary is delivered and grill waits.
Question format
End each active grill turn with exactly one numbered question. Give concrete options with trade-offs. Give a recommendation and rationale.
Probe vague words, overloaded terms, and hidden assumptions. Use concrete scenarios to test concept boundaries and edge cases. When code contradicts the user's claim, quote both the claim and the code behavior.
Open-question accounting
Use - [ ] for unresolved questions.
Use - [+] for questions added by the latest answer.
Use - [u] for questions updated by the latest answer.
Use - [x] for questions resolved by the latest answer.
Use - [-] for questions removed by the latest answer.
Show - [u], - [x], and - [-] for one turn only.
After that turn, keep an updated question visible as - [ ].
After that turn, remove resolved and removed questions from the visible list.
Add questions only when the latest answer directly causes them. When the user extends scope instead of answering, account for the new question before pursuing it. Do not add speculative questions to the open-question list. Put speculative follow-up questions in the final summary instead.
Include a cause note for each - [+], - [u], and - [-] change.
Use this format:
- [+] Policy action names
Cause: Your answer introduced `ALLOW` and `DENY`, changing the prior action names.
Preserve agent judgment when the user reframes or invalidates a question. Show the resulting list change instead of following a rigid reframe algorithm.
When the user catches incorrect open-question handling, briefly explain the error. Then show the corrected open-question list.
Turn endings
While grill is active, every assistant turn must end with exactly one of these items:
- the next numbered question
- the final decision summary
- a plain statement that grill is paused because the user changed tasks
If the turn includes a correction, file edit, apology, or explanation, still end with the next numbered question.