AGENT SKILL

using-grillspec

Use when starting any conversation, and before acting on any request that involves choosing between options - routes work through grilling, specs, plans and execution so the agent never guesses what the user would decide.

License 许可
MIT
所属仓库
forbxpg/grillspec
最近更新
2026-10-21
Stars
★ 0
Forks
0

01 前言 Front Matter

name
using-grillspec
description
Use when starting any conversation, and before acting on any request that involves choosing between options - routes work through grilling, specs, plans and execution so the agent never guesses what the user would decide.

02 资源文件 5

03 SKILL.md 正文

Using grillspec

grillspec spends questions up front so you never have to guess later. A choice between options is a decision, and decisions belong to the user.

The rule

Before you act on a request — before clarifying questions, before exploring code — decide which row applies:

The request Do
has no choice between options: a typo, a rename, formatting, a fix whose cause and remedy are both certain Just do it.
has a decision but is small: one function or file, reversible, no new interface Ask in chat: name the decision, give your recommendation, ask. One question, no skill, no files.
adds or changes a feature, module, interface, schema, infrastructure, or anything hard to reverse Run the full pipeline below.

Unsure between two rows? Take the lower one. One extra question costs less than one rework.

Debugging, test-driven development and verification skills (grillspec-debugging, grillspec-tdd, grillspec-verification) keep their own triggers whatever the row.

The pipeline

grill → spec (+ memory) → plan → execute → archive

Phase Skill Leaves behind
grill grillspec-grill, with grillspec-council when the council is on a decision log (Open / Settled), confirmed by the user
spec grillspec-specs, then grillspec-memory an approved spec; approved memory entries
plan grillspec-plans an approved plan
execute grillspec-subagent-execution or grillspec-inline-execution (the plan's executor), with grillspec-tdd, grillspec-debugging, grillspec-verification, then grillspec-finishing code and green tests; every task ticked
archive grillspec-archive the change merged into the living specs, or marked implemented

If a skill named here is not installed yet, tell the user which phase you are in and follow the phase by hand.

No .grillspec/config.toml in the project and the request is in row 2 or 3? Then offer the grillspec-init skill once; if the user declines, ask in chat without files. Row-1 work never needs it.

Where you are

Each session starts with a line like grillspec: phase=grill change=2026-10-17-login next=continue grilling: 3 open questions. When unsure, run grillspec status, using the command on the grillspec CLI: line of the session-start context (or uvx grillspec where no plugin is installed). Phases: idle, grill, spec, plan, execute, done.

Confirming a grill and approving a spec or a plan come only from an explicit reply from the user — never inferred from silence, from a "looks fine so far", or from your own review. An explicit "no questions" or "just do it" is consent to proceed on your own judgment. Before you change any file, write the decisions you are taking for the user as a visible list headed "Assumptions" — in your reply, not only in your reasoning; one line each, with what to change if it is wrong. Then do the work, and repeat the list in your final message. "I'm in a hurry" alone is not that consent: still ask the one decision that matters most, with your recommendation.

Facts and decisions

Look facts up yourself: files, git history, configuration, documentation, a CLI's --help. Asking the user what you could have checked wastes their time. Put every decision to the user: what depends on their product, taste or risk is theirs.

Asking

Each decision you put to the user carries your recommendation, first and marked (Recommended), with the real trade-off of each alternative. Use the harness's multiple-choice tool when it has one; otherwise write numbered options in plain text and accept a number or free text as the answer.

Language

Reply in the language of the user's latest message. Write specs and plans in the project's artifact language (spec.artifacts_language; auto means the language of the project's existing docs, or of the conversation in an empty repository). Keep OpenSpec keywords in English: SHALL, MUST, WHEN, THEN, AND, and the section headers.

Other workflow plugins

If Superpowers or another workflow router is also active, tell the user once to keep only one of them enabled. Until they do, follow grillspec for planning and for its phase gates.

Phase gates

While the active change is grilled, specced or planned, a hook blocks edits outside its artifacts. A blocked edit means: finish the phase first. Never work around a gate with another tool or a shell trick; if the user wants to skip it, they run grillspec gate pause in their own terminal.

Session memory

Your session-start context holds a grillspec memory: brief: where the last session stopped. If it is missing (GigaCode), run grillspec recall --brief once. Write a note the moment you learn something: grillspec note discovery|blocker|todo -- "<one line>". Before the session ends, write grillspec note handoff -- "done: …; next: …; open: …". Look further back with grillspec recall <words>; close a blocker or todo that no longer stands with grillspec recall --resolve <id>.

Red flags

Thought Reality
"This is obvious, I'll just build it." Obvious is where unchecked assumptions live. Check the rule table.
"I'll pick a sensible default." Picking between options is a decision. Ask.
"Let me explore the code first, then ask." Decide the row first; looking things up is part of the grill.
"The user is in a hurry." Ask the one decision that matters — unless they said no questions: then list your assumptions first and proceed.
"They approved the idea, so the spec is approved." Approval is an explicit reply to the written artifact.
"I'll leave a note todo and add the test later." A note records; it never defers a test or a gate. Write the test now — or, if the user waived it, say plainly the code is untested and leave the task unticked.

Harness tools

Skills name actions — ask the user, dispatch a subagent, track tasks, run a command. How each harness spells them:

  • Claude Code: references/claude-code-tools.md
  • Cursor: references/cursor-tools.md
  • Codex: references/codex-tools.md
  • Antigravity: references/antigravity-tools.md
  • GigaCode: references/gigacode-tools.md

04 同仓库更多 Skills