Plan with Sub-Agents Skill

These guidelines apply automatically whenever plan mode is active. Plan mode implies complexity, so always design for parallelism and minimal context.

Core Principle: Use Sub-Agents Generously

Sub-agents are cheap and effective. When in doubt, spawn an agent. Benefits:

• Parallel execution - Multiple agents work simultaneously
• Reduced context - Each agent only sees what it needs
• Fault isolation - One agent's failure doesn't block others
• Cleaner commits - Each agent commits its own work

Default to spawning agents for:

• Any exploration or research (use Explore agent)
• Designing approaches (use Plan agent)
• Independent file modifications (use general-purpose agents in parallel)
• Error-prone steps that involve noisy logs (e.g. running tests)
• Shell operations (use Bash agent)

When Sub-Agents Apply

• Any task in plan mode (plan mode = complexity assumed)
• Tasks with 4+ distinct steps
• Work that touches multiple unrelated files
• Tasks where exploration and execution can be separated
• Long-running operations that benefit from parallel execution

Phase Design Principles

1. Break by File Independence

Group work by which files are touched. Phases that modify different files can run in parallel.

Phase A: modifies src/api/server.py, src/api/models.py
Phase B: modifies tests/test_api.py, tests/conftest.py
Phase C: modifies README.md, docs/API.md
→ All three can run in parallel

2. Identify Sequential Dependencies

Some phases MUST run in order:

• Phase that creates a file → Phase that references it
• Phase that changes API → Phase that updates tests for that API
• Phase that installs deps → Phase that uses those deps

3. Minimize Context per Agent

Each sub-agent should receive ONLY the context it needs:

• Specific file paths to modify
• Exact changes to make
• Commit message to use

Avoid passing full project history or unrelated requirements.

Agent Type Selection

Execution Prompt Template

When spawning sub-agents, use this structure:

**Task:** [One sentence describing the goal]

**Files to modify:**
- path/to/file1.py - [what to change]
- path/to/file2.py - [what to change]

**Specific changes:**
1. [Concrete change 1]
2. [Concrete change 2]

**Commit when done:**
"[type]: [description]"

Example Plan Structure

markdown
1## Phase 0: Setup (sequential)
2- Create directory structure
3→ COMMIT: "chore: Add project scaffolding"
4
5## Phase 1: Research (sequential)
6- Explore codebase to understand patterns
7- No commit (exploration only)
8
9## Phases 2, 3, 4 (PARALLEL)
10Agent A - Phase 2: Backend changes
11- Modify src/api/*.py
12→ COMMIT: "feat: Add new endpoint"
13
14Agent B - Phase 3: Frontend changes
15- Modify src/ui/*.tsx
16→ COMMIT: "feat: Add UI component"
17
18Agent C - Phase 4: Documentation
19- Modify docs/*.md
20→ COMMIT: "docs: Document new feature"
21
22## Phase 5: Integration (sequential, after 2-4)
23- Update tests that depend on all changes
24→ COMMIT: "test: Add integration tests"
25
26## Phase 6: Validation
27- Run tests, linting, type checks
28→ COMMIT only if fixes needed

Parallelization Checklist

Before marking phases as parallel, verify:

• No shared files between parallel phases
• No phase reads output from another parallel phase
• Each phase can commit independently
• Merge conflicts are impossible

Anti-Patterns

Don't: Create one giant phase with all changes Do: Split by logical units that can be tested independently

Don't: Pass entire CLAUDE.md to every sub-agent Do: Extract only relevant sections for each task

Don't: Have parallel phases commit to the same file Do: Sequence phases that touch shared files

Commit Strategy

Each phase should commit its own work immediately when done. This provides:

• Clear git history
• Easy rollback if one phase fails
• Parallel agents don't block each other

Use conventional commit prefixes:

• feat: - New feature
• fix: - Bug fix
• chore: - Maintenance
• docs: - Documentation
• test: - Tests
• refactor: - Code restructuring