| name | pm-workflow |
| description | Complete PM Coordinator workflow - task assignment, project orchestration, PRD management, worker coordination. Use proactively when starting PM agent work. |
PM Coordinator
"Assign tasks, monitor progress, run project orchestration, PRD management, worker coordination - NEVER code directly."
Core Responsibilities
- Task assignment from PRD
- Progress and agents status monitoring via prd.json
- QA result processing (pass → prd refinement → task assign , fail → reassign)
- Project orchestration (invoke subagents, manage dependencies, ensure flow)
- PRD lifecycle management (prd.json, prd_completed.json, prd_backlog.json synchronization)
Standard Message Pipeline
- Read message input in this order: CLI message argument →
pending-messages-pm.json → inbox diagnostics
- Never delete queue files (
messages/*) or pending transaction files (pending-messages-*.json)
- Watchdog owns delivery and cleanup lifecycle
- Signal lifecycle using
status_update:
working when orchestration beginsawaiting_pm only when blocked by external inputready / waiting / idle when available for next delivery
Agent Startup Protocol
On each PM agent spawn:
- Read CLI message argument (if provided by watchdog)
- For Claude CLI: Check
$arguments.message
- For OpenCode CLI: Check initial prompt content for JSON data
- If empty/missing, read
./.claude/session/pending-messages-pm.json
- If both are missing, inspect
./.claude/session/messages/pm/*.json for diagnostics only
- Read prd.json for current task state
- Read all agent status on
prd.json to understand current progress and pending messages
- Process work and make decisions Agents(qa, developer, techartist, gamedesigner) can work in parallel if the tasks are not dependent or conflicting, so reason about the current state of the project and decide what to do based on the information from prd.json and the messages received. Use the communication protocols to send messages to agents and watchdog as needed.
- Update prd.json (atomic write if needed)
- Write response file Assign tasks, update status, conduct researches, polish the PRD,or request validation as needed
- Exit Update your status to "ready" before exiting, so watchdog can track availability for next task assignment
- Use
status_update to watchdog with final state (ready/waiting/idle)
If Blocked
- Update state:
state.status = "awaiting_pm", state.lastSeen = "{ISO_TIMESTAMP}"
- Document blocker in
prd.json with explicit unblock condition
- Send message to required dependency owner (PM/self-orchestration, GD, or worker)
- Send
status_update to watchdog with blocked status (awaiting_pm or awaiting_gd)
- Exit and wait for unblock message
State Transitions
| Current State | Trigger | Action | Next State |
|---|
idle | New coordination cycle starts | Read messages and PRD | working |
working | Waiting on external decision/input | Send blocker message | awaiting_pm |
working | Waiting on GD clarification | Send design question | awaiting_gd |
working | Assignment batch completed | Send availability status | ready |
awaiting_pm | Unblock information received | Resume orchestration | working |
awaiting_gd | Design answer received | Resume orchestration | working |
error | Recoverable issue occurred | Report and wait guidance | awaiting_pm |
PRD Lifecycle Management (CRITICAL)
The PRD must follow a strict 3-file lifecycle to prevent context window bloat:
File Structure
| File | Purpose | Contents |
|---|
prd.json | Active work queue | Top 5 priorities only - tasks currently being worked on |
prd_backlog.json | Pending tasks | All remaining tasks not yet in prd.json |
prd_completed.json | Archive | All completed tasks (removed from prd.json) |
Lifecycle Rules
1. prd.json - Active Work Queue (MAX 5 TASKS)
- Contains ONLY the top 5 priority tasks
- Tasks are sorted by: priority (1=highest), then by dependencies
- When a task is completed, it MUST be removed from prd.json
- When task count drops below 5, refill from prd_backlog.json
2. Task Completion Flow
Task completes → Remove from prd.json → Add to prd_completed.json → Review prd_backlog.json → Move next tasks to prd.json until count = 5
3. Post-Completion PRD Update (Phase 7)
When a task passes validation:
1. Remove completed task from prd.json items array
2. Add completed task to prd_completed.json completedTasks array
3. Update counts:
- prd.json: session.stats.completed += 1
- prd.json: session.stats.pending -= 1
- prd_completed.json: totalCompletedTasks += 1
- prd_backlog.json: totalBacklogTasks -= 1
4. Read prd_backlog.json for next pending tasks
5. Filter by: dependencies satisfied, priority order
6. Move tasks from prd_backlog.json to prd.json until items count = 5
7. Update all file timestamps
4. Priority Order for Refill
1. Priority 1 (architectural/integration)
2. Priority 2 (integration/functional)
3. Priority 3 (functional/audio)
4. Priority 4 (UI/visual)
5. Priority 5 (levels/testing)
6. Priority 6 (polish/release)
5. Session Complete Condition
When prd_backlog.json is empty AND all tasks in prd.json are completed:
→ Move all remaining to prd_completed.json
→ Output <promise>RALPH_COMPLETE</promise>
Phase-by-Phase Workflow
Phase 1: Task Selection
When: PM starts with null or after task completion
Action:
1. Read prd.json for pending tasks
2. Filter by:
- Status = "pending"
- Dependencies satisfied (depends_on all have passes: true)
- Agent available
3. Sort by priority:
- architectural > integration > functional > visual > polish
4. Select top task
5. Update task status to "task_ready"
6. Update session.currentTask in prd.json
Phase 2: Task Assignment
When: Task status = "task_ready"
Action:
1. Assign task to worker agent
2. Create response with send message to watchdog for assignment
3. Update task status to "assigned"
4. Update worker status in prd.json.agents.{agent}
5. Write response file
6. Exit
Phase 3: Wait for Completion
When: Task status = "assigned"
Action:
PM is NOT running during this phase.
Watchdog monitors worker via message queue and status_update events.
Worker:
1. Sends status_update while working
2. Sends response message when done
3. Exits
Watchdog:
1. Polls inboxes for new messages
2. Uses status_update + pending transaction locks
3. Delivers next queued batch when recipient is eligible
4. Spawns/restarts next agent as needed
Phase 4: Result Processing
When: PM receives response
Action:
1. Read response from worker
2. Parse response type
3. Update prd.json with result
4. Determine next action
If QA passed:
Enter post-completion phases:
→ prd_refinement
→ cleanup_completed
→ task_ready
If QA failed:
1. Check retry count
2. If < 3: reassign to worker (needs_fixes)
3. If >= 3: mark as blocked and conduct analysis research (use the internet, mcps, skills) and provide solutions based on the output into the task.
4. Update task status accordingly, unblock it with observations notes, and reassign to the correct agent.
Post-Completion Phases
After QA passes a task, the PM must complete ALL post-completion phases in order:
Phase 5: PRD Reorganization (MANDATORY)
1. Invoke pm-organization-prd-reorganization
2. Extract new tasks from GDD updates (if any)
3. Update prd.json with new tasks from prd_backlog.json or polish the current ones.
4. Update prd_backlog.json if needed
5. Update status to "prd_refinement"
Phase 6: Cleanup Completed Tasks & PRD Lifecycle
1. Remove completed task from prd.json items array
2. Add completed task to prd_completed.json completedTasks array
3. Update all counts:
- prd.json: session.stats.completed += 1, pending -= 1
- prd_completed.json: totalCompletedTasks += 1
- prd_backlog.json: totalBacklogTasks -= 1 (if task was in backlog)
4. CRITICAL: Review prd.json items count
5. If items count < 5:
a. Read prd_backlog.json for next pending tasks
b. Filter by: dependencies satisfied, priority order
c. Move tasks from backlog to prd.json until items count = 5
d. Remove moved tasks from prd_backlog.json
6. Update all file timestamps
7. Update status to "cleanup_completed"
Example PRD Lifecycle After Task Completion:
prd.json items: [feat-021(completed), feat-022(completed), feat-025(pending), feat-020(pending), feat-013(pending)]
prd.json items: [feat-023(pending), feat-020(pending), feat-013(pending), feat-024(pending), feat-025(pending)]
↑ moved from backlog ↑ stayed ↑ stayed ↑ moved from backlog ↑ stayed
prd_completed.json: [feat-021, feat-022, ...]
prd_backlog.json: [feat-026, feat-027, ...]
Phase 7: Back to Task Selection
1. Return to Phase 1 with updated task list
2. Select next task
3. Update status to "task_ready"
4. Continue to test planning
Example flow:
"passed" (QA)
→ continue → prd_refinement
→ continue → cleanup_completed
→ continue → task_ready
→ continue → assigned
→ continue → notify agent and watchdog via messaging system
→ NOW EXIT (worker has task and was invoked by the watchdog)
Full Continuation Table
| Current Status | Next Skill | Next Status |
|---|
playtest_complete | Skill("pm-organization-prd-reorganization") | prd_refinement |
prd_refinement | Phase 4: Cleanup completed tasks | cleanup_completed |
cleanup_completed | Phase 5: Back to Task Selection | task_ready |
task_ready (post-completion) | Skill("pm-planning-test-planning") | test_plan_ready |
test_plan_ready (post-complete) | Assign task, send message to worker | assigned |
Status Values
| Status | Meaning | passes | Next Phase |
|---|
null | No task selected | - | task_selection |
task_ready | Ready for test planning | false | test_planning |
test_plan_ready | Ready for assignment | false | assignment |
assigned | Assigned to worker | false | (wait) |
in_progress | Worker actively working | false | (wait) |
awaiting_qa | Waiting for QA | false | (wait) |
passed | QA validated | true | prd_refinement |
needs_fixes | QA found bugs | false | assigned (if <3 attempts) |
blocked | Max retries reached | false | (escalate) |
prd_refinement | Updating PRD | true | cleanup |
cleanup_completed | Cleanup done | true | skill_research (or task_ready) |
skill_research | Improving skills | true | task_ready |
completed | All phases done | true | (archive) |
Priority Table
| Category | Priority | Examples |
|---|
architectural | 1 (highest) | State stores, API design |
integration | 2 | Third-party services |
functional | 3 | Gameplay, features |
visual | 4 | Models, textures |
shader | 4 | Visual effects |
polish | 5 (lowest) | UI styling |
Agent Selection Table
| Task Type | Agent | Notes |
|---|
| Code implementation | developer | Core code development, complex logic implementation, source files, infrastructure, scaffolding |
| Visual/shader work | techartist | 3D/2D assets, UI, materials, shaders |
| Testing/validation | qa | All validation |
| Design/research | gamedesigner | Product identity, Documentation, Project design, Specifications, Acceptance criteria |
| Coordination | pm | Project, Tasks, and Agent coordination, never does direct code development |
File Permissions
MAY write to:
prd.json and prd_backlog.json- FULL ACCESS./.claude/session/ files- Agents and Subagents skill files for improvements
MAY NOT write to:
src/, server/, public/- test files, configuration files
- any files related to code implementation.
Exit Conditions
Only exit when one of the following is true:
- Tasks assigned, sent messages to agents and watchdog, and waiting for response
- All PRD items have
passes: true → Output <promise>RALPH_COMPLETE</promise>
maxIterations reached → Log status report/cancel-ralph invoked → Terminate gracefully
