Batch

Batch Skill

When to activate

• Applying the same change across 10+ files (rename, refactor, migration)
• Running a large codebase audit (security scan, dependency check, test coverage)
• Generating boilerplate for many modules in parallel
• Any task where the work can be divided into independent, non-overlapping units

When NOT to use

• Tasks with sequential dependencies (step B requires step A's output)
• Changes to a single file or a small number of related files
• Tasks that require shared context across all units (use a single agent instead)
• When you need to review and approve each change before the next starts

Instructions

The batch pattern

Standard Claude Code works sequentially: one task → one agent → one session. Batch mode breaks a large task into N independent units and processes them in parallel — each unit runs as a separate background agent in an isolated git worktree, making its changes and opening a PR.

Large task
    │
    ├── Unit 1 → worktree-1 → branch-1 → PR #1
    ├── Unit 2 → worktree-2 → branch-2 → PR #2
    ├── Unit 3 → worktree-3 → branch-3 → PR #3
    └── Unit N → worktree-N → branch-N → PR #N

Activation prompt

/batch

Task: [describe the full task]
Files/scope: [list files or glob patterns, or describe the scope]

Claude will:

1. Research phase — read the codebase to understand patterns and scope
2. Decomposition — break the task into 5–30 independent units
3. Plan review — present the breakdown and wait for your approval
4. Execution — spawn one background agent per unit in an isolated worktree
5. PRs — each agent commits its changes and opens a PR against main

Decomposition rules Claude follows

• Each unit must be independent — no shared state, no inter-unit dependencies
• Each unit must be completable in one agent session (~15–30 min of work)
• Each unit must have a clear success criterion (tests pass, lint passes)
• Units are sized to be reviewable in one PR (prefer small PRs over large ones)

Good batch tasks

# Rename a function used across the codebase
/batch
Task: Rename `getUserById` to `findUserById` everywhere it's used.
Scope: src/**/*.ts, tests/**/*.ts

# Add type annotations to all Python modules
/batch
Task: Add full type annotations (PEP 484) to all functions in the services layer.
Scope: src/services/*.py

# Migrate API calls to new SDK
/batch
Task: Migrate all uses of the old `stripe.charges.create()` to `stripe.paymentIntents.create()`.
Scope: src/billing/**

# Security audit
/batch
Task: Audit every endpoint handler for missing authentication middleware.
Scope: routes/**/*.ts
Report findings per file — do not make changes.

Monitoring progress

While agents run in the background, monitor with:

# Check worktree status
git worktree list

# Check open PRs
gh pr list --label batch

# See which agents are still running
claude agents

Merging results

Once PRs are open:

1. Review each PR independently — they're small by design
2. Merge in any order (they're independent)
3. Clean up worktrees after all PRs are merged:

git worktree prune

When a unit fails

If one agent's PR fails tests:

• The other agents continue — failures don't cascade
• Review the failing PR, fix manually or re-run that unit
• Use git worktree remove worktree-N to clean up and restart

Example

Task: Add JSDoc comments to all exported functions in a 40-file TypeScript library.

Claude's decomposition:

Unit 1: src/auth/*.ts (6 files, ~15 functions)
Unit 2: src/billing/*.ts (5 files, ~12 functions)
Unit 3: src/api/users/*.ts (4 files, ~18 functions)
...
Unit 8: src/utils/*.ts (3 files, ~8 functions)

After approval: 8 background agents start in parallel. Each opens a PR titled docs(jsdoc): add JSDoc to [module name]. Total time: ~20 minutes instead of ~2.5 hours sequential.