AGENT SKILL

grilling

Shared interview engine for /brainstorm, /prd, and grill-me. Asks only the open frontier of a question set, looks up facts instead of asking them, and returns a question count. Engine only; not invoked on its own.

License 许可
MIT
最近更新
2026-11-12
Stars
★ 95
Forks
18

01 前言 Front Matter

name
grilling
description
Shared interview engine for /brainstorm, /prd, and grill-me. Asks only the open frontier of a question set, looks up facts instead of asking them, and returns a question count. Engine only; not invoked on its own.
allowed-tools
Read, Grep, Glob, AskUserQuestion, Task, Bash(bash .claude/skills/grilling/frontier.sh:*), Bash(bash .claude/skills/_shared/journal.sh:*), Bash
user-invocable
false

02 资源文件 2

03 SKILL.md 正文

Grilling Engine

Engine skill. Caller skills (/brainstorm, /prd, grill-me) load this file and run it with an input contract. Users do not run it directly. If a user types /grilling, tell them to use /brainstorm, /prd, or grill-me instead.

Pattern source: mattpocock/skills skills/productivity/grilling (design tree, frontier rounds, facts are the agent's job). HQ changes the default: one question per AskUserQuestion call, with rounds available only on request.

Input contract

Input Required Meaning
question_set yes Path to a question-set file. Format: .claude/skills/_shared/questions/FORMAT.md.
known_facts no Map of question id to answer, already settled by the caller (from the brief, a brainstorm, the board, or earlier answers). These ids are never asked.
mode no one-at-a-time (default) or rounds.
max_questions no Upper bound on questions put to the user. Default: no limit. When reached, remaining decisions take their recommended option and are logged as assumed.

mode: rounds is set only when the user explicitly asks for rounds in this session or the caller received a --rounds argument. Never infer it from the size of the set, the user's tone, or time pressure.

Helper

.claude/skills/grilling/frontier.sh parses the set and computes the frontier. Use it rather than computing dependencies by hand.

bash .claude/skills/grilling/frontier.sh frontier <question_set> --known <ids> --answered <ids>
bash .claude/skills/grilling/frontier.sh count <question_set> --known <ids>

frontier prints the open decision ids in file order. count validates the set and prints the count object for a full run. Exit 2 means the set is malformed or has a dependency cycle; stop and report the error to the caller.

Procedure

  1. Validate. Run frontier.sh count with the known ids. On exit 2, return the error to the caller without asking anything.

  2. Resolve facts. For every kind: fact question whose depends_on are resolved, find the answer yourself: caller research, the company board, repo files, or a Task subagent lookup. Never put a fact question to the user. A pending lookup blocks only the questions that depend on it; keep asking the rest of the frontier.

  3. Compute the frontier. Run frontier.sh frontier with known ids plus everything answered so far. A question whose dependency is still open waits for a later pass.

  4. Ask.

    • one-at-a-time (default): one AskUserQuestion call per frontier question. Offer 2 to 4 options from the question's options. When research or known facts support the recommended option, put it first and append (Recommended) to its label. Wait for the answer before the next call.

    • rounds: present the whole frontier as one numbered block, then wait for the reply.

      Q1 - <question text>
      Options: A) <opt> B) <opt> C) <opt>
      Recommended: A) <opt>
      
      Q2 - ...
      

      Word each question so "yes" accepts the recommendation. Accept shorthand replies: yes or y (accept all recommendations), 1A 2C, 2: <free text>, or all recommended except 3B. Questions not mentioned in a reply take the recommendation only when the reply was yes; otherwise ask about them again in the next round.

  5. Record. After each answer, append it to the active journal:

    bash .claude/skills/_shared/journal.sh append <project_dir> decisions "<question id>: <answer>"
    

    Facts found in step 2 go under findings with their id and source.

  6. Repeat steps 2 to 5 until the frontier is empty or max_questions is reached.

  7. Return the answers map and the count object to the caller.

Headless

When CLAUDE_HEADLESS=1, do not call AskUserQuestion. Take the recommended option for each decision, log it under decisions with the suffix (assumed, headless), and count it under asked.

Count object

{"asked": 7, "skipped_known": 3, "skipped_fact": 2,
 "by_tier": {"strategic": 1, "architecture": 3, "quality": 3}}
  • asked: decisions put to the user (or assumed under headless or max_questions).
  • skipped_known: ids supplied in known_facts.
  • skipped_fact: fact questions resolved by lookup.
  • by_tier: asked split by tier.

Test: bash core/scripts/tests/grilling-count.test.sh.

04 同仓库更多 Skills