grilling
Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- grilling
- description
- Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.
02 资源文件 1
- 参考文档agents/openai.yaml113 B
03 SKILL.md 正文
Interview the user relentlessly until you reach a shared understanding. Map this as a design tree: every decision branches into the decisions that hang off it.
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 haven't heard yet. Ask the frontier in one round: number each question and give your recommended answer. Then wait for the user's answers before the next round.
When a message from the user arrives during a round without answering its questions (a remark, another request), answer it, then end your reply with the open round given again in full, in its format (the table, then each question still unanswered with your recommended answer), not with a pointer to a message above: questions already answered do not come back. That round given again stays the same round: no new question joins it, and the bound of 4 questions and about 400 words holds for it, the answer to the message not counted. The user had to ask "give me all the questions again" after such a message.
A round holds 4 questions at most, and about 400 words: rounds of 750 to 1,050 words went unread. When the frontier holds more, ask first the questions whose answers unblock the most others; the rest of the frontier waits for the next round.
Format a round like so: a two-column table first, the question in a few words and your recommended answer in one line, then the detail of each question under it.
| Question | Recommended |
|---|---|
| 1. <question in a few words> | <your recommended answer, one line> |
| 2. <question in a few words> | <your recommended answer, one line> |
❓ **Q1** - **<question title>**: <the fact it rests on, one sentence>. <question body, a few lines at most, including multiple choices>
➡️ <your recommended answer>
---
❓ **Q2** - **<question title>**: <the fact it rests on, one sentence>. <question body, a few lines at most, including multiple choices>
➡️ <your recommended answer>
Word each question so "yes" accepts your recommended answer. Each question reads on its own: it restates in one or two sentences the fact or the finding it rests on, instead of pointing to a report the user may not have read. In a round, that fact holds in one sentence, under the table, at the head of its question; the body of a question stays within a few lines.
Each round the user answers reshapes the tree: settled decisions push the frontier outward and unblock questions that depended on them. Recompute the frontier and ask the next round. A question whose answer depends on another question still open in this round belongs to a later round, not this one. When the user imposes a mechanism (a tool, a format, an approach), the next round's first question is about the mechanism itself; its settings wait for the round after, since they fall if the mechanism changes.
Finding facts is your job, never the user's. When a frontier question needs a fact from the environment (filesystem, tools, etc.), dispatch a sub-agent to find it; don't ask the user for anything you could look up yourself. Don't block on it: a running exploration is an unsettled prerequisite, so only the questions downstream of it wait for the sub-agent to report; ask the rest of the frontier now, within the round's bound of 4 questions. The decisions are the user's: put each to them and wait.
The session is done when the frontier is empty: every branch of the design tree visited, nothing left silently assumed. Do not act on it until the user confirms you have reached a shared understanding.
04 同仓库更多 Skills
reserve
Skill d'essai de la sonde, réservé à l'utilisateur. Répond par un mot fixe, sans rien faire d'autre.
recap
Récap de l'usage Claude Code en chat, d'après ccusage et les relevés de requêtes de stats-board — la discussion en cours (coût, tokens, cache, durée, temps actif, attente de ta réponse, temps API, requêtes et la plus chère, modèles, spécialistes) et la journée (coût, tokens, temps actif, temps cumulé et temps API de ce poste, par projet et par modèle, discussions actives, bloc de 5 h et son rythme, autres postes publiés dans claude-usage), ou une mission entière (manager, workers et spécialistes, retrouvés par le début de leur titre ou par les worktrees groupe-* d'un projet, avec leur temps actif sans double compte et leur temps cumulé). À utiliser sur « /stats-board:recap », « récap de l'usage », « combien m'a coûté cette discussion », « combien j'ai consommé aujourd'hui », « coût de la journée », « récap de la mission », « combien a coûté la mission », « récap du parcours », « combien a coûté le parcours ». Pas pour ouvrir le panneau de statistiques, qui passe par /ccstats.
install
Installe et initialise les plugins de la marketplace mathis-plugins choisis par l'utilisateur - choix par groupes, dépendances et prérequis, installation, réglages (Asana, partage des usages, modèle de session-tools, barre d'état) et bloc de règles dans le CLAUDE.md global ou du projet, montré avant écriture. Relançable pour ajouter ou retirer des plugins.
version
Bump the project version with SemVer - detect the version files (package.json, pyproject.toml, Cargo.toml, Expo app config, Claude Code plugin manifests and their marketplace entry...), propose the bump from the commits since the last release, update every file consistently; handles Expo native vs OTA releases.
sync
Bring the current branch up to date with its remote, or with the default branch when asked, without losing local work - fetch, rebase or merge, autostash.
push
Finish a piece of work and push it - commit what is left, run the project's checks, confirm the end-of-session files are up to date, sync with the remote and push without force.