prd-taskmaster（珠三角工头）

PRD Generator for TaskMaster v3.0

Smart PRD generation with deterministic operations handled by script.py. AI handles judgment (questions, content, decisions); script handles mechanics.

Script location: ~/.claude/skills/prd-taskmaster/script.py All script commands output JSON.

When to Use

Activate when user says: PRD, product requirements, taskmaster, task-driven development. Do NOT activate for: API docs, test specs, project timelines, PDF creation.

Core Principles

• Quality Over Speed: Planning is 95% of the work
• Taskmaster Required: Blocks if not detected
• Engineer-Focused: Technical depth, code examples, architecture
• Validation-Driven: 13 automated checks via script
• User Testing Checkpoints: Every 5 tasks

Workflow (12 Steps)

Step 1: Preflight & Resume Detection

python3 ~/.claude/skills/prd-taskmaster/script.py preflight

Returns JSON: has_taskmaster, prd_path, task_count, tasks_completed, tasks_pending, taskmaster_method, has_claude_md, has_crash_state, crash_state.

If has_crash_state is true: Present resume options to user:

1. Continue from last subtask
2. Restart current task
3. Resume from last checkpoint
4. Start fresh

Then proceed to Step 2.

Step 2: Detect Existing PRD

Use preflight JSON: if prd_path is not null and task_count > 0, an existing PRD is found.

If existing PRD found, use AskUserQuestion:

• Execute tasks from existing PRD (skip to Step 11)
• Update/refine existing PRD (edit and re-parse)
• Create new PRD (replace - backup first via script.py backup-prd --input <path>)
• Review existing PRD (display summary, then exit)

If no PRD found: Proceed to Step 3.

Step 3: Detect Taskmaster

Use preflight JSON field taskmaster_method: mcp, cli, or none.

If none: Block and show installation instructions:

• Option 1 (recommended): Install MCP Task-Master-AI
• Option 2: npm install -g task-master-ai
• Wait for user to install and confirm, then re-run: script.py detect-taskmaster

No proceeding without taskmaster detected.

Step 4: Discovery Questions

Ask detailed questions to build comprehensive PRD. Use AskUserQuestion for structured input.

Essential (5):

1. What problem does this solve? (user pain point, business impact)
2. Who is the target user/audience?
3. What is the proposed solution or feature?
4. What are the key success metrics?
5. What constraints exist? (technical, timeline, resources)

Technical (4): 6. Existing codebase or greenfield? 7. Tech stack? 8. Integration requirements? 9. Performance/scale requirements?

TaskMaster-specific (3): 10. Used taskmaster before? 11. Estimated complexity? (simple/typical/complex) 12. Timeline expectations?

Open-ended (1): 13. Anything else? (edge cases, constraints, context)

Smart defaults: If user provides minimal answers, use best guesses and document assumptions.

Step 5: Initialize Taskmaster

Only if .taskmaster/ doesn't exist (check preflight has_taskmaster).

python3 ~/.claude/skills/prd-taskmaster/script.py init-taskmaster --method <cli|mcp>

For MCP: use the returned params to call mcp__task-master-ai__initialize_project. For CLI: script runs taskmaster init directly.

Step 6: Generate PRD

Load template:

python3 ~/.claude/skills/prd-taskmaster/script.py load-template --type <comprehensive|minimal>

Returns JSON with content field containing the template.

AI judgment: Fill template with user's answers from Step 4:

• Replace placeholders with actual content
• Expand examples with project-specific details
• Add technical depth based on discovery answers

Write completed PRD to .taskmaster/docs/prd.md.

Step 7: Validate PRD Quality

python3 ~/.claude/skills/prd-taskmaster/script.py validate-prd --input .taskmaster/docs/prd.md

Returns JSON: score, max_score, grade, checks (13 items), warnings.

Grading: EXCELLENT (91%+), GOOD (83-90%), ACCEPTABLE (75-82%), NEEDS_WORK (<75%).

AI judgment: If warnings exist, offer user three options:

1. Proceed with current PRD
2. Auto-fix warnings
3. Review and fix manually

If grade is NEEDS_WORK, strongly recommend fixing before proceeding.

Step 8: Parse & Expand Tasks

Calculate task count:

python3 ~/.claude/skills/prd-taskmaster/script.py calc-tasks --requirements <count>

Returns recommended task count.

For MCP:

mcp__task-master-ai__parse_prd: input=".taskmaster/docs/prd.md", numTasks=<recommended>, research=true
mcp__task-master-ai__expand_all: research=true

For CLI:

taskmaster parse-prd --input .taskmaster/docs/prd.md --research --num-tasks <recommended>
taskmaster expand-all --research

Step 9: Insert User Test Tasks

python3 ~/.claude/skills/prd-taskmaster/script.py gen-test-tasks --total <task_count>

Returns array of USER-TEST task definitions with title, description, dependencies, template.

For each task in the array:

• MCP: mcp__task-master-ai__add_task with title, description, details=template, dependencies, priority=high
• CLI: taskmaster add-task --title="..." --description="..." --dependencies="..." --priority=high

Step 10: Setup Tracking Scripts

python3 ~/.claude/skills/prd-taskmaster/script.py gen-scripts --output-dir .taskmaster/scripts

Creates 5 scripts: track-time.py, rollback.sh, learn-accuracy.py, security-audit.py, execution-state.py.

Step 10.5: Generate CLAUDE.md

Pre-check: Use Glob to check if ./CLAUDE.md exists. If it exists, skip.

If generating:

1. Load template: script.py load-template won't work here -- use Read tool on ~/.claude/skills/prd-taskmaster/templates/CLAUDE.md.template
2. AI judgment: Replace placeholders with project-specific values from discovery:

- {{PROJECT_NAME}}, {{TECH_STACK}}, {{ARCHITECTURE_OVERVIEW}} - {{KEY_DEPENDENCIES}}, {{TESTING_FRAMEWORK}}, {{DEV_ENVIRONMENT}}, {{TEST_COMMAND}}

1. Write to ./CLAUDE.md
2. Ask if user uses Codex -- if yes and no codex.md, write identical copy

Step 11: Choose Next Action

Use AskUserQuestion:

Question: "PRD and tasks ready. How to proceed?"

• Show TaskMaster Commands (default): Display command reference, then exit skill
• Autonomous Execution: Ask follow-up for execution mode

If Autonomous Execution selected, ask execution mode:

• Sequential to Checkpoint (recommended): Tasks one-by-one until next USER-TEST
• Parallel to Checkpoint: Independent tasks in parallel until USER-TEST
• Full Autonomous: All tasks parallel, skip user validation
• Manual Control: User decides each task

AI judgment: Recommend mode based on context:

• First-time/critical: Sequential
• Experienced/non-critical: Parallel
• Trusted/time-critical: Full Autonomous
• Complex/learning: Manual

Step 12: Summary & Start

If Handoff: Display PRD location, task counts, key requirements, validation score, task phases, user test checkpoints, and TaskMaster commands. Then exit skill.

If Autonomous: Display same summary plus execution mode, then begin execution using the selected mode's rules.

Execution Mode Rules

All Modes Include

• DateTime tracking: python3.taskmaster/scripts/track-time.py start|complete <task_id> [subtask_id]
• Progress logging: python3 ~/.claude/skills/prd-taskmaster/script.py log-progress --task-id <id> --title "..." --duration "..." --subtasks "..." --tests "..." --issues "..."
• Git policy: Branch per task (task-{id}-{slug}), sub-branch per subtask, merge to main with checkpoint tag
• Rollback: If user says "rollback to task X", run bash.taskmaster/scripts/rollback.sh X
• State tracking: python3.taskmaster/scripts/execution-state.py start|complete|checkpoint <task_id>

Sequential to Checkpoint

Execute tasks one-by-one. For each task:

1. Start time tracking
2. Create feature branch
3. For each subtask: create sub-branch, implement, test, commit, merge to task branch
4. Complete time tracking
5. Log progress
6. Merge to main, create checkpoint tag
7. Stop at next USER-TEST for user validation

Parallel to Checkpoint

Same as sequential but launch up to 3 concurrent independent tasks. Handle merge conflicts automatically. Stop at USER-TEST.

Full Autonomous

Maximum parallelization (up to 5 concurrent). Auto-complete USER-TEST tasks. Only stop when ALL tasks complete.

Manual Control

Wait for user commands: "next task", "task {id}", "status", "parallel {id1,id2}".

Tips

• More detail in discovery = better PRD
• Quantify goals: not "improve UX" but "increase NPS from 45 to 60"
• USER-TEST checkpoints catch issues early
• Git checkpoints allow easy rollback
• Use script.py validate-prd at any time to re-check PRD quality