| name | breakdown |
| description | Break down a PRD or CRD into self-contained implementation tasks for LLM execution. Use when you have a PRD/CRD file and want to generate executable task files for autonomous implementation. |
| context | fork |
| allowed-tools | Read Glob Grep Write Skill Bash |
| model | claude-sonnet-4-5 |
/breakdown - PRD/CRD Task Breakdown
You are orchestrating the breakdown of a PRD (Product Requirements Document) or CRD (Change Request Document) into implementation tasks optimized for autonomous LLM execution.
Arguments
<input-path>: Path to PRD index.md or CRD document (required)
--layer <N>: Only generate tasks for specific layer (optional, 0-4)
--review-only: Only run review pass on existing tasks (optional)
--output-dir <path>: Target directory for greenfield projects (overrides default)
--project-path <path>: Existing project path for brownfield/CRD (overrides PRD value)
--auto-setup: Automatically execute Layer 0 tasks after generation (greenfield only)
Input Format Detection
The skill automatically detects whether the input is a PRD or CRD:
| Root Element | Format | Description |
|---|
<prd> | PRD | Full product requirements (greenfield or brownfield) |
<crd> | CRD | Change request for existing project |
CRD handling:
- CRDs are always brownfield (no Layer 0)
- CRDs require
--project-path or a project with PROJECT.md
- CRDs use
<impact-analysis> to scope task generation
- CRDs typically produce fewer tasks (focused changes)
Output Location
Tasks are written to: docs/tasks/<prd-slug>/
For greenfield with --output-dir:
- Layer 0 tasks reference the output-dir as target
- Other layer tasks are still saved to docs/tasks/
Workflow
Execute these phases in order:
Phase 1: Validate Input
- Read the input file at the provided path
- Detect format by checking root element (
<prd> or <crd>)
- Verify valid XML structure
- Extract the slug from
<meta><slug> tag
- Create output directory:
docs/tasks/{slug}/
- If directory exists, check for existing
.done markers to resume
For CRD input:
- Require
--project-path argument (CRDs always target existing projects)
- Verify PROJECT.md exists at
{project-path}/PROJECT.md
- Load PROJECT.md context for use in task generation
Phase 2: Analyze Input
For PRD:
Invoke the breakdown-analyze-prd skill with the full PRD content.
Pass the PRD XML content and request structured extraction of:
- All features with priorities
- Tech stack with versions
- Implied data models
- Implied API endpoints
- Implied frontend components
- External dependencies
- Template path if specified
For CRD:
Extract directly from CRD structure:
- Requirements from
<requirements> section
- Acceptance criteria from
<acceptance-criteria> section
- Affected files from
<impact-analysis><affected-files>
- Affected features from
<impact-analysis><affected-features>
- Tech stack from PROJECT.md context
- Related existing features from
<context><related-features>
The CRD already contains impact analysis, so less inference is needed.
Save the analysis to docs/tasks/{slug}/analysis.json
Phase 3: Plan Layers
For PRD:
Invoke the breakdown-plan-layers skill with the analysis JSON.
Request organization into 4-5 layers:
- 0-setup: Template copy, git init, environment (greenfield only)
- 1-foundation: Database models, migrations, base config
- 2-backend: API endpoints, services, business logic
- 3-frontend: React components, state management, routing
- 4-integration: Wiring, E2E flows, polish
For CRD:
Layer planning is simpler based on <impact-analysis>:
- If
<affected-schemas> has changes: Include Layer 1 (foundation)
- If
<affected-apis> has changes: Include Layer 2 (backend)
- If frontend files in
<affected-files>: Include Layer 3 (frontend)
- Always include Layer 4 (integration) for wiring changes together
CRD typically produces 2-3 layers, not 5.
Save the layer plan to docs/tasks/{slug}/layer_plan.json
Phase 4: Generate Tasks (Per Layer)
Determine layers to process based on input type:
- PRD Greenfield:
[0-setup, 1-foundation, 2-backend, 3-frontend, 4-integration]
- PRD Brownfield:
[1-foundation, 2-backend, 3-frontend, 4-integration] (skip Layer 0)
- CRD: Only layers identified in Phase 3 based on impact analysis
For each layer in order:
-
Check completion: If {layer}/.done exists, skip this layer
-
Create directory: docs/tasks/{slug}/{layer}/
-
Batch tasks: Split layer tasks into batches of max 5 tasks each
- If layer has ≤5 tasks: single batch
- If layer has 6-10 tasks: 2 batches
- If layer has 11-15 tasks: 3 batches
- Example: 14 tasks → batches of [5, 5, 4]
-
For each batch, with retry loop (max 3 attempts per batch):
for batch in batches:
for attempt in [1, 2, 3]:
# Generate
invoke breakdown-generate-tasks with:
- Layer name
- Task batch (subset of layer tasks)
- PRD analysis
- Template path (if any)
- Output directory: docs/tasks/{slug}/{layer}/
- Previous review feedback (if retry)
# Review
invoke breakdown-review-tasks with:
- Path to generated task files
- Layer name
# Handle result
if review.verdict == "PASSED":
break # Move to next batch
elif attempt < 3:
# Extract failures for next attempt
feedback = {
"attempt": attempt + 1,
"previous_failures": review.critical_issues
}
else:
# Max retries exceeded
Report: "Batch failed after 3 attempts. Manual intervention required."
Report: review.critical_issues
Exit without creating .done
-
Mark layer complete: After ALL batches pass, create {layer}/.done marker
Phase 5: Finalize
- Generate
docs/tasks/{slug}/manifest.json containing:
- PRD slug and name
- Generation timestamp
- Layer completion status
- Task inventory with IDs and file paths
- Report completion summary:
- Total tasks generated
- Tasks per layer
- Any review failures requiring attention
Task File Format
Each task file follows this XML structure (see references/task-format-spec.md for full spec):
<task>
<meta>
<id>L1-001</id>
<name>Task Name</name>
<layer>1-foundation</layer>
<priority>1</priority>
</meta>
<context>...</context>
<dependencies>...</dependencies>
<objective>...</objective>
<requirements>...</requirements>
<test-requirements>...</test-requirements>
<files-to-create>...</files-to-create>
<verification>...</verification>
<exports>...</exports>
</task>
Critical Constraints
- Self-contained tasks: Each task file MUST contain ALL information needed for implementation. No external lookups.
- Small context: Tasks are designed for ~50k token context models (Haiku, GLM 4.5-4.7)
- Interface contracts: Dependencies use type signatures, not full code
- Max 3 files per task: Keep scope manageable
- TDD approach: Test requirements come before implementation
- Explicit verification: Every task has runnable verification commands
Error Handling
- If PRD file not found: Report error, exit
- If PRD invalid XML: Report parsing error with details, exit
- If skill invocation fails: Report which phase failed, suggest retry
- If review fails: Do NOT mark layer complete, report specific issues
Example Usage
Greenfield Project
/breakdown docs/prd/voice-prd-generator/index.md --output-dir /path/to/new-project
Output:
Analyzing PRD: voice-prd-generator
Detected: greenfield project (Python + FastAPI template)
Target directory: /path/to/new-project
Saved analysis to: docs/tasks/voice-prd-generator/analysis.json
Planning layers...
Saved layer plan to: docs/tasks/voice-prd-generator/layer_plan.json
Generating Layer 0 (Setup)...
Batch 1/1: [L0-001, L0-002, L0-003, L0-004]
- L0-001-copy-template.xml
- L0-002-init-git-env.xml
- L0-003-configure-database.xml
- L0-004-verify-setup.xml
Reviewing batch... PASSED
Created: docs/tasks/voice-prd-generator/0-setup/.done
Generating Layer 1 (Foundation)...
Batch 1/1: [L1-001, L1-002, L1-003, L1-004, L1-005]
- L1-001-project-model.xml
- L1-002-conversation-model.xml
- L1-003-message-model.xml
- L1-004-prddocument-model.xml
- L1-005-personaworkflow-model.xml
Reviewing batch... PASSED
Created: docs/tasks/voice-prd-generator/1-foundation/.done
Generating Layer 2 (Backend)...
Batch 1/3: [L2-001, L2-002, L2-003, L2-004, L2-005]
Reviewing batch... PASSED
Batch 2/3: [L2-006, L2-007, L2-008, L2-009, L2-010]
Reviewing batch... FAILED (attempt 1)
- L2-008: Contains placeholder 'TBD' for schema
Regenerating with feedback...
Reviewing batch... PASSED (attempt 2)
Batch 3/3: [L2-011, L2-012, L2-013, L2-014]
Reviewing batch... PASSED
Created: docs/tasks/voice-prd-generator/2-backend/.done
[continues for layers 3-4...]
Breakdown complete!
- Total tasks: 28
- 0-setup: 4 tasks
- 1-foundation: 5 tasks
- 2-backend: 14 tasks
- 3-frontend: 7 tasks
- 4-integration: 4 tasks
Brownfield Project (PRD)
/breakdown docs/prd/new-feature/index.md --project-path /existing/project
Output:
Analyzing PRD: new-feature
Detected: brownfield project
Existing project: /existing/project
Skipping Layer 0 (setup)
[continues with layers 1-4...]
CRD (Change Request)
/breakdown docs/crd/dark-mode-toggle.md --project-path /existing/project
Output:
Analyzing CRD: dark-mode-toggle
Detected: Change Request Document
Existing project: /existing/project
Loading PROJECT.md context...
Impact Analysis:
- Affected files: 4 (2 modify, 2 create)
- Affected features: settings
- Breaking changes: none
Planning layers from impact...
- Layer 2 (backend): 1 task (API endpoint)
- Layer 3 (frontend): 2 tasks (component, hook)
- Layer 4 (integration): 1 task (wiring)
Generating Layer 2 (Backend)...
Batch 1/1: [L2-001]
- L2-001-theme-settings-api.xml
Reviewing batch... PASSED
Created: docs/tasks/dark-mode-toggle/2-backend/.done
Generating Layer 3 (Frontend)...
Batch 1/1: [L3-001, L3-002]
- L3-001-theme-toggle-component.xml
- L3-002-use-theme-hook.xml
Reviewing batch... PASSED
Created: docs/tasks/dark-mode-toggle/3-frontend/.done
Generating Layer 4 (Integration)...
Batch 1/1: [L4-001]
- L4-001-wire-theme-toggle.xml
Reviewing batch... PASSED
Created: docs/tasks/dark-mode-toggle/4-integration/.done
Breakdown complete!
- Total tasks: 4
- 2-backend: 1 task
- 3-frontend: 2 tasks
- 4-integration: 1 task
To execute:
/execute docs/tasks/dark-mode-toggle/ --project-path /existing/project