メニュー
Co-author plan.md INTERACTIVELY with a human through an interview (grill-with-docs style) for a needs-plan Issue. Agents never author a Plan alone — this Skill is the human-input gate. Invoked by /plan <id>.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Implement a ready Issue inside its own git worktree, building only within existing architecture. Cord-pull to needs-plan if hidden architecture is discovered. Use when advancing an Issue out of issues/ready/.
Re-validate an Issue against current main and perform the real git merge to main. Merges are serialized (one at a time). Use after the Validation Pipeline passes (auto-merge) or after a human merge-review approves.
Second pass over an Issue — explore the codebase, produce a candidate file list and deterministic signals, and FINALIZE Difficulty. Use for any Issue not fast-pathed as trivial by Triage.
First pass over an inbox Issue — assign Type and a provisional Difficulty from the Issue text ALONE (no code exploration). Use when advancing an Issue out of issues/inbox/.
Run the pre-merge Validation Pipeline on an Issue's worktree — intent-judge, then tests, then deterministic checks — each with a bounded fix-and-retry loop. Use for an Issue in validating/, and again at merge time for re-validation.
| name | plan |
| description | Co-author plan.md INTERACTIVELY with a human through an interview (grill-with-docs style) for a needs-plan Issue. Agents never author a Plan alone — this Skill is the human-input gate. Invoked by /plan <id>. |
A Plan is co-authored interactively with a human. You may not write it alone — if
no human is in the loop, stop and say so. The Plan IS how human architectural input enters
the system; "plan required" and "human approval required" are the same event
(DESIGN.md §4, ADR 0001).
grill-with-docs skillissues/needs-plan/<id>/ (has a ## Scoping section).plan.md in the Issue folder.plan.md sections (DESIGN.md §4)---
id: <id>
plan-author: human+agent
planned: <date>
---
# Intent
Outcome the Issue achieves. The intent-judge checks the diff delivers this.
# Approach
The chosen design and which existing patterns/abstractions it builds on. For architectural
Issues, the human-supplied structure lives here. Freeform headers can be used to describe extra sections. Be very explicit about the required implementation so another agent can take this plan and blindly implement it.
# Files expected to change
Explicit list. SUPERSEDES Scoping's candidate list (you may add/remove during the
interview). Bounds the diff the intent-judge will accept.
# Out of scope (MANDATORY)
What this Issue deliberately does NOT do — the "minimal thing" boundary. The intent-judge's
over-delivery check flags scope creep against this. Do not omit it.
Return ready once plan.md is complete and the human is satisfied. /plan then moves the
Issue needs-plan/ → ready/, where it rejoins the agent implementation queue.