AGENT SKILL

task-planner

When decomposing a feature or spec into implementable subtasks, estimating scope, or orchestrating multi-step work across subagents.

License 许可
MIT
最近更新
2026-11-07
Stars
★ 13
Forks
3

01 前言 Front Matter

name
task-planner
license
MIT
description
When decomposing a feature or spec into implementable subtasks, estimating scope, or orchestrating multi-step work across subagents.

03 SKILL.md 正文

Task Planner Standards

Task Management

  1. Plan First: Write plan with checkable items before starting.
  2. Verify Plan: Check in with the user before executing.
  3. Track Progress: Mark items complete as you go.
  4. Explain Changes: High-level summary at each step.
  5. Document Results: Add review section when complete.
  6. Capture Lessons: Update .memory/LESSONS.md after corrections.

Task Quality

  • Tasks must be clear, specific, and actionable.
  • Complex tasks must be broken into atomic, trackable subtasks.
  • Dependencies between tasks must be explicitly declared.
  • Security-related tasks must undergo mandatory review.
  • Escalate ambiguous or contradictory tasks for clarification.

Blueprint Check (if configured)

If Blueprint Path is set in .memory/CONFIG.md and the task involves shared business logic:

  • [ ] Read the relevant spec(s) in [BLUEPRINT_PATH]/specs/ before planning.
  • [ ] Check [BLUEPRINT_PATH]/parity/PARITY_MATRIX.md for current platform status.
  • [ ] Include in the plan: "Requires Knowledge Bus entry" if the change affects cross-platform contracts.
  • [ ] Note any spec gaps — if the spec is missing or incomplete, plan a spec update as part of the task.

Skip this section if no blueprint is configured or the task is project-specific.

Change Tracking

  • All rule or config changes must be documented in CHANGELOG.md.
  • Record source, timestamp, and rationale when modifying shared assets.
  • A rollback plan must be defined for every major change.

Subagent Strategy

When a task benefits from fresh context, parallel execution, or isolated scope, use the subagent-dispatch skill. It provides prompt templates for implementer, reviewer, and researcher subagents with explicit scope boundaries and result validation.

When to dispatch:

  • Subtasks are independent and can proceed in parallel
  • Parallel execution provides meaningful time savings
  • Fresh context helps (no accumulated conversation noise)
  • Scope of each subtask is clear and self-contained

When NOT to dispatch:

  • Iterative feedback from the user is needed between steps
  • The task is simpler to do directly in the current session
  • Subtasks need full conversation history or shared state
  • Scope is unclear or likely to shift during execution

04 同仓库更多 Skills