AgentSkillsCN

orchestration

涵盖 UI/UX 设计、无障碍访问,以及设计系统构建。内容包括设计原则、结合 WCAG 对比度要求的色彩理论、字体比例尺、8pt 行间距网格、响应式断点、WCAG 2.1 无障碍合规标准、动效设计,以及设计系统治理。在设计界面、选择配色,或当用户询问“这个界面是否具备无障碍特性?”或“我该采用怎样的行间距?”时,可作为参考。

SKILL.md
--- frontmatter
name: orchestration
description: >-
  Coordinates multi-agent work with PM-style patterns. Covers agent delegation, council deliberation
  workflows, session file management, Linear issue integration, and product planning. Use when
  managing complex tasks across agents, running decision councils, or when the user asks "how do
  I break down this work?" or "which agent should handle this?"

PM Orchestration

Comprehensive patterns for PM-style orchestration: coordinating multi-agent work, managing sessions, running councils, delegating to specialized agents, and integrating with Linear.

Philosophy

PM is the orchestrator, not the implementer.

The PM agent:

  1. Creates issues and session files for tracking
  2. Breaks down work into delegable tasks
  3. Spawns specialized agents for implementation
  4. Coordinates outcomes and updates external systems
  5. Never implements code, tests, or documentation directly

Every release should be complete, polished, and delightful - no MVPs or quick hacks.

Quick Reference

TaskAction
Multi-step workCreate session file, spawn agents
Complex decisionConvene council (5-7 agents, odd)
Linear updateCheckboxes, no emoji, no local paths
Feature planningDefine appetite, shape before building
Agent selectionMatch domain expertise to task
Stuck on taskCheck circuit breaker, consider reshaping
Pre-compactionSpawn context-archiver to preserve state
Low-priority workSpawn background-runner with run_in_background
New feature workflowResearch → Architecture → PRD → Specs → Tasks
Create specificationUse /specs with requirement reference
Break down specUse /tasks to generate atomic work items
Task-coupled sessionUse /implement TASK-XXX for traceability

Topics

TopicReferenceUse When
Product Developmentreferences/product-development.mdFollowing Research → Vision → Architecture → Specs workflow
Specificationsreferences/specs.mdCreating specs, shaping work, defining test conditions
Local Tasksreferences/local-tasks.mdManaging tasks locally or with Linear backend
Agent Delegationreferences/delegation.mdChoosing agents, spawning subagents, decision trees
Parallel Agentsreferences/parallel-agents.mdDispatching independent work concurrently
Subagent Developmentreferences/subagent-development.mdDelegating to specialized agents
Background Agentsreferences/background-agents.mdRunning non-interactive work in background
Council Workflowreferences/councils.mdConvening councils for complex decisions
Session Managementreferences/sessions.mdCreating sessions, handoffs, validation
Session Resumereferences/session-resume.mdResuming sessions, checkpoints, context recovery
Context Managementreferences/context-management.mdUsing /clear, /compact, managing context limits
Linear Integrationreferences/linear.mdUpdating Linear issues, magic words, status conventions
Product Planningreferences/planning.mdShape Up methodology, setting appetite, roadmaps

Configuration

This skill uses paths from .agents/config.json:

json
{
  "sessions": {
    "directory": ".agents/sessions",
    "councils_directory": ".agents/councils"
  },
  "linear": {
    "workspace": "your-workspace-slug",
    "project": { "id": "...", "name": "..." },
    "default_team": "Platform"
  }
}

Available Scripts

ScriptUsageDescription
new-session.shnew-session.sh <description> [linear-issue]Generate session file
new-council.shnew-council.sh <topic> <session> <agents...>Generate council file
validate-session.pyvalidate-session.py <file>Validate session format
validate-council.pyvalidate-council.py <file>Validate council format
validate-roadmap.pyvalidate-roadmap.py <file>Validate roadmap format
get-config.pyget-config.py [key.path]Read config values
suggest-team.pysuggest-team.py "task desc"Suggest Linear team
check-linear-format.pycheck-linear-format.py <file>Validate Linear text
format-progress.shformat-progress.sh "Done" -- "Todo"Format progress update
extract-magic-words.shextract-magic-words.sh HEAD~10..HEADExtract Linear refs

Three-Phase Workflow

BEFORE (Planning)

  • Create/check external issue (Linear, GitHub)
  • Create session file for internal coordination
  • Break down into tasks
  • Identify which agents to spawn
  • Get user approval before spawning

DURING (Execution)

  • Spawn specialized agents (never implement directly)
  • Track progress in session file
  • Post progress to external issue
  • Convene councils for uncertain decisions

AFTER (Completion)

  • Run code review with backend/frontend devs (if significant changes)
  • Run QA testing and security checks
  • Update external issue to Done
  • Ensure knowledge captured in permanent locations
  • Archive session file (set status, archived_at, archived_by, move to .agents/sessions/archive/, update .agents/ links)

When to Use PM Orchestration

Use for:

  • Feature implementation
  • Non-trivial bug fixes
  • Refactoring
  • Infrastructure changes
  • Multi-agent coordination

Skip for:

  • Typo fixes
  • Direct questions
  • Quick clarifications
  • Single-file trivial changes

Critical Rules

Sessions

  • Filename: YYYYMMDD-HHMMSS-description.md
  • Required fields: title, status, created, last_updated, current_task
  • Archive fields: archived_at, archived_by (required when archived)
  • Required sections: Context, Current State, Next Steps
  • Archive when complete (set status, archived_at, archived_by, move to .agents/sessions/archive/, update .agents/ links)

Context Preservation (PreCompact)

  • PreCompact hook identifies sessions modified in last 60 minutes
  • Spawn context-archiver agent when prompted by hook
  • Agent generates ## Resumption Prompt section for seamless continuation
  • After compaction, read session's Resumption Prompt to pick up where you left off

Transcript Archival (Post-Compaction)

  • Claude Code provides transcript path after /compact or /clear
  • Copy transcript to .agents/transcripts/ (create directory if needed)
  • Add filename to session's transcripts: array in frontmatter
  • Preserves full audit trail for context recovery

Councils

  • Always odd number: 5 or 7 agents
  • Councils advise, users decide
  • PM coordinates but doesn't vote
  • Spawn all agents in parallel
  • Document decision after user approval

Linear

  • Checkboxes only (- [x]), no emoji
  • Outcome-focused, self-contained
  • No local file references
  • Use issue ID only (Linear auto-expands)
  • Magic words in commit body, not subject

Planning (Shape Up)

  • Appetite over estimates (decide time, flex scope)
  • Shape before building (boundaries, not tasks)
  • Circuit breakers at 50% to reassess
  • Now/Next/Later buckets, no version numbers
  • Clear acceptance criteria (Given/When/Then)
  • No backlogs - bet or let go