本文へ移動
cccskills
無料GitHub で公開

pm-as-solo-orchestrator

Use when planning a delivery request as the only orchestration agent - one board-writing subtask, backlog first, one approval question, then move approved tasks to todo

インストール方法を見る

含まれるファイル(1)

  • SKILL.md3.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

PM as Solo Orchestrator

Overview

When product-manager is the only enabled orchestration agent, the plan you write for a request is not a multi-agent handoff — it's one subtask that writes the board, followed by columns doing the rest of the work automatically.

Core principle: One subtask creates the tasks; the board, not the plan, drives every stage after that.

Rules

  • All subtasks are assigned to product-manager — never to system-architect, security-agent, backend-developer, frontend-developer, mobile-developer, data-scientist, game-developer, or qa-agent. Developer work goes through create_board_task, not an orchestration subtask.
  • Multiple parallel PM subtasks are fine only when every one is read-only (e.g. web research + list_board_tasks concurrently). At most one subtask per parallel group may write to the board; a second board writer in the same group rejects the whole plan. Use depends_on to sequence when a later subtask needs an earlier one's result.
  • Delivery stages (analiz, implementation, QA verification, pm_uat, approval) are board columns the task travels through afterwards — never plan a subtask per stage; that opens one board record per stage for a single piece of work and is rejected before the plan runs.
  • Analiz tasks skip approval — they go straight to todo (investigation is always safe to start).
  • Planner sets ready=false only when a product decision blocks creating any board task (max 3 questions, see stakeholder-intake's ask-vs-assume rule). Otherwise ready=true.

Standard flow — one subtask, in order

  1. Create the tasks in backlog (see implementation-task-spec, acceptance-criteria-gwt).
  2. Call ask_user, choice mode:
    • context: what you asked / what you're assuming (name each assumption so the stakeholder can correct it cheaply) / the task list with assignee and order.
    • options: "Start all (recommended)", "Start only some", "Change something", plus whatever else applies.
  3. On approval, move every approved task to todo with move_board_task. Order lives in blocked_by — never hold a task back in backlog to fake an order; a blocked task parks itself and starts automatically once its blocker lands.
  4. On partial approval, move only the selected tasks; leave the rest in backlog.
  5. On rejection or requested changes, update_board_task the affected fields, or delete_board_task + create_board_task for assignee/derived_from (creation-only fields) — and for task_type when update_board_task's schema does not list it.
  6. Report which tasks moved to todo (agents will pick them up) and which stay in backlog, and why.

This is ONE subtask calling these tools in order — not three subtasks. Splitting it hands the same board record to several agents that cannot see each other's writes.

If a blocking product decision remains, call ask_user within a subtask — never append questions as markdown in the final summary.

Common Mistakes

  • A subtask per delivery stage instead of one subtask that creates the tasks.
  • Holding implementation tasks in backlog "until analiz is done" instead of using blocked_by — the board already enforces that order.
  • Two parallel subtasks writing to the board in the same group.
  • Approval-question context that doesn't separate what you asked from what you assumed.

Red Flags

  • A plan with a subtask named "QA stage" or "pm_uat review" — those are columns, not subtasks.
  • More than one board-writing subtask in a parallel group.
  • A task left in backlog with a "Depends on: …" sentence instead of blocked_by.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use when writing acceptance criteria for a task - express each as an observable Given/When/Then that QA can execute, including negative cases

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

Use when the diff adds or changes an endpoint, resolver, RPC, job or query that takes an object id, a role check, a request binding or a tenant filter - BOLA/IDOR, function-level authorization, mass assignment and tenant scoping

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

Use on every UI change - semantic HTML, labels for controls, keyboard-navigable dialogs/menus, visible focus, and never color as the only signal

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

Use when a task changes any screen, form, dialog, menu or control - Lighthouse/axe scan of the changed screens, a keyboard walk, and the thresholds that fail a task

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

How to work a task returned with review, QA or UAT findings. Use when a task is in need_revision or PR review comments are in your context.

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

Use when deciding whether a request needs an analiz task before implementation - the conditions that require the architect's analysis versus going straight to implementation

日本語の概要は準備中です。原文の説明を表示しています。

makifbaysal/tasktrooper1122026年10月12日 更新

makifbaysal のスキルをすべて見る

このスキルの問題を報告する