/grill-me
Frontier/round-based plan interview. Triggers: grill me, interview me about this plan, shared understanding, walk the design tree. NOT one-question-at-a-time.
$ golems-cli skills install grill-meUpdated today
Replaces the v1 stub that pointed at a
/grillingskill which does not exist on this machine, and the retired one-question-per-turn method. Source: Matt Pocock, skills v1.2 (digestdocs.local/video-gems/2026-08-06-matt-pocock-skills-v1.2.0.md). One-question-per-turn was "incredibly frustrating and dead slow" once only easy yes/no questions remained.
Scope
The interview settles what is already decided, then asks ONLY the open independent decisions in numbered rounds with lettered options and one Recommended each; answer in shorthand (1A 2C).
The loop
- Settle first. Read the plan and the codebase. List what is already decided, with the line or file that decides it. Do not ask about anything on that list. If a question can be answered by exploring the codebase, explore the codebase instead of asking.
- Compute the frontier. The frontier is the set of open decisions that do not depend on another open decision. Only frontier items get asked this round. Dependent items wait.
- Ask one round. Number the questions (
1.,2., …). Give each lettered options (A,B,C) and mark exactly one Recommended with a one-line why. Three to six questions per round; never one, never twelve. - Read shorthand answers. The user answers
1A 2C 3Bor1A, 2: something else. Never answer for them. Never treat silence as a choice. If an answer is not one of the options, take it verbatim as the decision. - Recompute the frontier with the new decisions folded in. Dependent items that just became independent are the next round.
- Empty frontier = stop. Emit ≤6 frozen bullets — the decisions, each with the round it was made in — and wait for the user to say shared understanding is reached. Do not start implementing.
Rules Etan set
- Never answer for him. A Recommended option is a recommendation, not a default. Unanswered questions stay open; they do not resolve to Recommended.
- Options, not essays. A question with no options is a question you have not finished thinking about. Go back and finish.
- Voice-friendly. The format exists so a round can be answered by dictation in one breath
(
1A 2B 3A). Keep option labels short enough to say aloud. - Codebase before questions. Anything greppable is not a question.
Full SKILL.md source — includes LLM directives, anti-patterns, and technical instructions stripped from the Overview tab.
Replaces the v1 stub that pointed at a
/grillingskill which does not exist on this machine, and the retired one-question-per-turn method. Source: Matt Pocock, skills v1.2 (digestdocs.local/video-gems/2026-08-06-matt-pocock-skills-v1.2.0.md). One-question-per-turn was "incredibly frustrating and dead slow" once only easy yes/no questions remained.
Scope
The interview settles what is already decided, then asks ONLY the open independent decisions in numbered rounds with lettered options and one Recommended each; answer in shorthand (1A 2C).
The loop
- Settle first. Read the plan and the codebase. List what is already decided, with the line or file that decides it. Do not ask about anything on that list. If a question can be answered by exploring the codebase, explore the codebase instead of asking.
- Compute the frontier. The frontier is the set of open decisions that do not depend on another open decision. Only frontier items get asked this round. Dependent items wait.
- Ask one round. Number the questions (
1.,2., …). Give each lettered options (A,B,C) and mark exactly one Recommended with a one-line why. Three to six questions per round; never one, never twelve. - Read shorthand answers. The user answers
1A 2C 3Bor1A, 2: something else. Never answer for them. Never treat silence as a choice. If an answer is not one of the options, take it verbatim as the decision. - Recompute the frontier with the new decisions folded in. Dependent items that just became independent are the next round.
- Empty frontier = stop. Emit ≤6 frozen bullets — the decisions, each with the round it was made in — and wait for the user to say shared understanding is reached. Do not start implementing.
Rules Etan set
- Never answer for him. A Recommended option is a recommendation, not a default. Unanswered questions stay open; they do not resolve to Recommended.
- Options, not essays. A question with no options is a question you have not finished thinking about. Go back and finish.
- Voice-friendly. The format exists so a round can be answered by dictation in one breath
(
1A 2B 3A). Keep option labels short enough to say aloud. - Codebase before questions. Anything greppable is not a question.
Output shape per round
## Round N — <frontier size> open decisions
1. <decision, one line>
A. <option>
B. <option> ← Recommended: <one-line why>
C. <option>
2. …
Answer like: 1B 2A 3C
Termination shape
## Frozen (Round N reached an empty frontier)
- <decision> — R1
- <decision> — R2
…
Say "shared understanding" to close, or name what is still open.
Changelog entries are derived from eval runs and skill version updates. Full cascading changelog (Phase 4D) coming soon.
- +Initial release to Golems skill library