project-coordinator

Project Coordinator Skill

Launch project coordination for complex, uncertain tasks.

When to Use

• Tasks with high uncertainty—where the solution path is unclear, multiple retries are expected, or progress tends to get lost without active tracking
• Bug investigation, performance debugging
• New library/API exploration (docs vs reality gaps)
• Environment setup issues
• Work requiring multiple approach attempts
• Tasks where you keep coming back to 'what was I doing?'
• .claude/project-coordinator/ directory exists (ongoing project)

NOT for: Predictable, low-risk tasks that existing agents handle well.

When Invoked

1. Check .claude/project-coordinator/ for existing docs
2. purpose.md first: Read existing. If missing or unclear, read purpose-extraction skill and apply it to clarify with user.

Agent Teams mode (when CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is available)

Execute the following steps exactly in order. Do not skip, reorder, or improvise.

⚠️ Critical rules — violating ANY of these breaks the workflow:

• coordinator gets ONE task: the full project objective
• investigator gets ZERO tasks from the lead
• The lead MUST NOT split work into subtasks — coordinator handles all breakdown
• The lead MUST NOT assign tasks to investigator — coordinator does this via SendMessage

Step 1: TeamCreate

TeamCreate(
  team_name: "project-coordinator",
  description: "<1-line summary from purpose.md>"
)

Step 2: TaskCreate — ONE project task

Create exactly one task. This is the entire project objective, not a subtask.

TaskCreate(
  subject: "<user's task in imperative form>",
  description: "## Purpose\n<purpose.md content>\n\n## Task\n<user's original request>",
  activeForm: "<present continuous form of the task>"
)

Record the returned task ID for Step 4.

Step 3: Spawn teammates — two Task calls in a single message (parallel)

Task(
  subagent_type: "project-coordinator:coordinator",
  name: "coordinator",
  team_name: "project-coordinator",
  description: "Coordinate project investigation",
  prompt: "You are the coordinator for this team.\nYour teammate is \"investigator\" — send investigation tasks via SendMessage.\nCheck TaskList, claim your task, read purpose.md, create plan.md, then begin."
)

Task(
  subagent_type: "project-coordinator:investigator",
  name: "investigator",
  team_name: "project-coordinator",
  description: "Investigate on coordinator direction",
  prompt: "You are the investigator for this team.\nWait for coordinator to send you tasks via messages.\nDo NOT start working until coordinator gives you a specific task."
)

Step 4: TaskUpdate — assign to coordinator

TaskUpdate(
  taskId: <task ID from Step 2>,
  owner: "coordinator"
)

Do NOT set status: "in_progress" — coordinator does this when it claims the task.

Step 5: Lead is done

After Step 4, the lead's launch work is complete.

• Coordinator reads TaskList, claims the task, creates plan.md, delegates investigation steps to investigator via SendMessage
• Investigator waits idle until coordinator sends a message, then begins work
• Coordinator enters monitoring loop (Sleep → check-in → evaluate → act)
• Coordinator messages the lead when: step complete, plan revision needed, or 5 "NO" threshold hit

The lead should wait for coordinator's reports. Do not poll or micromanage. User can message either teammate directly (Shift+Up/Down in terminal).

Subagent mode (when Agent Teams is not available)

Read ${CLAUDE_PLUGIN_ROOT}/agents/coordinator.md and apply its principles in this session.

For investigation tasks, call investigator via Task tool:

Task tool:
  subagent_type: "project-coordinator:investigator"
  prompt: "## Context\n[Purpose summary]\n\n## Step\n[Current step]\n\n## Task\n[Investigation details]"

⚠️ CRITICAL: Never just mention "delegate to investigator". Always use Task tool explicitly.