/grill-me
Design-tree interview. Triggers: grill me, interview my plan, stress-test this design. NOT for implementation or one-question-at-a-time interviews.
$ golems-cli skills install grill-meUpdated 3 weeks ago
Ported and adapted from Matt Pocock's
skills/productivity/grilling/SKILL.mdandskills/productivity/grill-me/SKILL.mdat upstream commitc55ee46073ed923f86ce59a5eb3b6d895095d1b7. Upstream is Copyright (c) 2026 Matt Pocock and MIT licensed; see LICENSE.upstream.Local shape: the upstream
grillingbody is folded into this user-invoked skill so a one-line wrapper does not add a second always-on description to the skill roster.Deliberate local divergence from upstream: retain Etan's voice-friendly lettered options,
1A 2Cshorthand, exactly oneRecommendedlabel per question, and the final <=6 frozen bullets. A recommendation is never a default.
Interview the user relentlessly until you reach a shared understanding. Map the subject as a design tree: every decision branches into the decisions that hang off it.
Work the whole frontier
Work the tree in rounds. The frontier is every decision whose prerequisites are already settled: the questions you can ask now without guessing at answers you have not heard yet.
Ask the whole frontier in one round. Do not impose a fixed question-count cap. Number every question, give it short lettered options, and mark exactly one option Recommended with a one-line reason. Then wait for the user's answers before the next round.
Each round reshapes the tree. Fold in the user's decisions, recompute the frontier, and ask the next round. A question whose answer depends on another open question belongs to a later round, not the current one.
Find facts without blocking the round
Finding facts is your job, never the user's. Read the plan and codebase first. When a frontier question needs a fact from the filesystem, tools, or another available source, dispatch a sub-agent to find it instead of asking the user.
Do not block the whole round on that exploration. Treat the running exploration as an unsettled prerequisite: hold only its downstream questions and ask the rest of the frontier now. Fold the fact into the tree when the sub-agent reports.
Facts belong to the agent. Decisions belong to the user: put every decision to them and wait.
Round format
Keep labels short enough to answer by voice in one breath:
Full SKILL.md source — includes LLM directives, anti-patterns, and technical instructions stripped from the Overview tab.
Ported and adapted from Matt Pocock's
skills/productivity/grilling/SKILL.mdandskills/productivity/grill-me/SKILL.mdat upstream commitc55ee46073ed923f86ce59a5eb3b6d895095d1b7. Upstream is Copyright (c) 2026 Matt Pocock and MIT licensed; see LICENSE.upstream.Local shape: the upstream
grillingbody is folded into this user-invoked skill so a one-line wrapper does not add a second always-on description to the skill roster.Deliberate local divergence from upstream: retain Etan's voice-friendly lettered options,
1A 2Cshorthand, exactly oneRecommendedlabel per question, and the final <=6 frozen bullets. A recommendation is never a default.
Interview the user relentlessly until you reach a shared understanding. Map the subject as a design tree: every decision branches into the decisions that hang off it.
Work the whole frontier
Work the tree in rounds. The frontier is every decision whose prerequisites are already settled: the questions you can ask now without guessing at answers you have not heard yet.
Ask the whole frontier in one round. Do not impose a fixed question-count cap. Number every question, give it short lettered options, and mark exactly one option Recommended with a one-line reason. Then wait for the user's answers before the next round.
Each round reshapes the tree. Fold in the user's decisions, recompute the frontier, and ask the next round. A question whose answer depends on another open question belongs to a later round, not the current one.
Find facts without blocking the round
Finding facts is your job, never the user's. Read the plan and codebase first. When a frontier question needs a fact from the filesystem, tools, or another available source, dispatch a sub-agent to find it instead of asking the user.
Do not block the whole round on that exploration. Treat the running exploration as an unsettled prerequisite: hold only its downstream questions and ask the rest of the frontier now. Fold the fact into the tree when the sub-agent reports.
Facts belong to the agent. Decisions belong to the user: put every decision to them and wait.
Round format
Keep labels short enough to answer by voice in one breath:
## Round N
1. <question title>: <question body>
A. <option>
B. <option> (Recommended: <one-line reason>)
C. <option>
---
2. <question title>: <question body>
A. <option> (Recommended: <one-line reason>)
B. <option>
C. <option>
Answer like: 1B 2A, or 1B 2: <answer outside the options>Separate every pair of questions in a round with a horizontal rule.
Read shorthand such as 1A 2C as the user's decisions. If an answer is outside the options,
take it verbatim. Never answer for him. Recommended is advice, not a default. Never treat
silence as a choice; unanswered questions remain open.
Stop only at shared understanding
The session is ready to close only when the frontier is empty: every branch of the design tree was visited and nothing remains silently assumed. Emit <=6 frozen bullets summarizing the settled decisions and the round that settled each one. Then wait.
Do not act on the plan until the user explicitly confirms shared understanding. If the user names something still open, put it back into the tree and continue.
Not yet evaluated on current models
Results will appear here once runs with a recorded model are published.
Changelog entries are derived from skill version updates. Full cascading changelog (Phase 4D) coming soon.
- +Initial release to Golems skill library
- +3 eval scenarios defined