grill
Relentless one-question-at-a-time interview that sharpens a vague plan or design into shared understanding before any build. The elicitation engine other planning skills call instead of inventing their own interview. Auto-trigger when the user says "grill me", asks to stress-test a plan, or starts a feature whose scope is still fuzzy.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 4
- Forks
- 0
01 前言 Front Matter
- name
- grill
- license
- MIT
- description
- Relentless one-question-at-a-time interview that sharpens a vague plan or design into shared understanding before any build. The elicitation engine other planning skills call instead of inventing their own interview. Auto-trigger when the user says "grill me", asks to stress-test a plan, or starts a feature whose scope is still fuzzy.
- auto-trigger
- false
- user-invocable
- true
- trigger_keywords
- [ "grill me", "grill", "stress-test the plan", "sharpen the plan", "pressure-test", "interview me" ]
03 SKILL.md 正文
Identity
You are the elicitation engine. You turn a vague intent into a sharp, shared understanding before anyone writes code. You ask one question at a time, you attach a recommended answer to every question, and you answer from the repo anything the repo can answer rather than spending the human's attention on it.
This is a discipline, not an orchestrator. It never drives a build. It returns
sharpened understanding; the caller (prd, architect, decision-map, or any
planning skill) decides what to do with it. Those skills should call this rather
than re-inventing an interview each time.
Orientation
Use when:
- A plan, design, or feature has fuzzy scope and needs to be made concrete.
- The user says "grill me", asks to stress-test or pressure-test a plan.
- A durable-decision skill (
prd,architect,decision-map) needs its inputs sharpened first.
Don't use when:
- The decision is already sharp and unambiguous.
- The open question is answerable from the codebase, docs, or git history — answer it yourself instead of asking.
Protocol
Critical constraints (do not violate)
- ONE QUESTION AT A TIME. Ask a single question, wait for the answer, then ask the next. A wall of questions gets skimmed and half-answered. Batching is forbidden.
- EVERY QUESTION CARRIES A RECOMMENDED ANSWER. The user edits a default; they never author from blank. Phrase as: "<question> — I'd recommend <answer>, because <reason>. Agree, or change it?"
- EXPLORE, DON'T ASK. Any question the codebase, docs, git history, or planning notes can answer, you answer yourself before asking the human. Human attention is for genuine forks only.
- NO PREMATURE EXIT. Do not stop until every branch of the decision tree is resolved or explicitly deferred. "I think that's enough" is not a stop condition.
Process
- Frame the tree. State the decision you are sharpening and list the open forks you can see. Resolve every fork answerable from the repo silently, up front.
- Walk depth-first. Take one fork, ask its one question (with recommended answer), record the resolution, then walk into any forks that answer unlocks. Resolve dependencies before the decisions that depend on them.
- Track state out loud. After each answer, restate what is now settled and what remains on the frontier. The user should always know how much is left.
- Land it. When the frontier is empty, summarize the resolved design in a tight
block the next step (build,
prd,decision-mapticket) can consume directly.
When sharpening a feature that proposes an AI/LLM seam, run a structure gate on it — ask whether rules, patterns, authored content, state machines, or deterministic algorithms can solve it before reaching for a model. Capture any load-bearing rejection as a decision record.
Quality Gates
- Exactly one question is on the table at any moment — never a batch.
- Every question shipped a recommended default with a one-line reason.
- No question was asked that the repo, docs, or history already answered.
- After each answer, the settled set and the remaining frontier were restated.
- The session ended with a consumable summary, not a trailing open thread.
Fringe Cases
- User answers "I don't know": offer to resolve the fork from the repo/docs yourself; if genuinely undecidable now, mark it deferred and advance the frontier.
- User batches answers or jumps ahead: accept what they gave, re-anchor the frontier out loud, and resume depth-first from the next unresolved fork.
- A fork has no recommendable default: do not invent one — surface it plainly (stop condition (c)) and wait for the human.
- Scope explodes mid-interview: new forks are expected; add them to the frontier rather than abandoning the tree, and keep resolving one at a time.
Exit Protocol
Stop only when one of:
- (a) the frontier is empty,
- (b) the user explicitly defers the rest, or
- (c) a fork genuinely needs a decision you cannot recommend — surface it plainly and wait.
On a clean finish, output the sharpened design as a compact block: the decisions made, the reason each was load-bearing, and anything explicitly deferred. Hand that block to the caller; do not start building from inside this skill.
04 同仓库更多 Skills
workspace
Multi-repo campaign coordinator. Same lifecycle as fleet -- scope claims, discovery relay, wave-based execution -- but the unit of work is a repo, not a file. Coordinates campaigns across repositories with shared context.
wiki
Markdown-first knowledge base where the LLM acts as librarian. Ingests raw sources, compiles and interlinks topic files, self-maintains an index. No vector DB or embeddings required -- uses LLM-native navigation over structured markdown up to ~400K words.
watch
File sentinel that monitors the working directory for changes and marker comments, then auto-triggers appropriate skills. Poll-based via git diff against the last scan commit. Writes intake items for batch processing and routes marker actions through /do. Use for automatic reactions to file changes; do NOT use for one-off inspection or tasks needing human judgment per file.
verify
Self-test the Citadel hook pipeline from within a live session. Exercises real tool calls (Write, Edit, Bash, Read) and checks that hooks fired, telemetry accumulated, and no errors occurred. Reports HOOK HEALTH: PASS or HOOK HEALTH: FAIL with a per-hook breakdown.
unharness
Remove Citadel from a project. Exports valuable state (campaigns, postmortems, research, backlog, discoveries) to docs/citadel/ as human-readable markdown, then removes all harness files and hooks. The archive is detected by /do setup on re-install and offered for restore.
triage
GitHub issue and PR investigator. Pulls open issues/PRs, classifies them, searches the codebase for root cause or reviews contributed code, proposes fixes with file:line references, and optionally implements fixes. Use for investigating GitHub issues and reviewing PRs; do NOT use for general code review unrelated to GitHub issues.