AGENT SKILL

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

License 许可
Apache-2.0
最近更新
2026-11-15
Stars
★ 122
Forks
18

01 前言 Front Matter

name
pm-as-solo-orchestrator
category
pm
description
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

03 SKILL.md 正文

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, 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 "human_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.

04 同仓库更多 Skills