Work Batch

Batch Execution Mode

Autonomous and sequential execution of user stories from a PRD file.

PRD file format

JSON (prd.json)

{
  "project": "project-name",
  "stories": [
    {
      "id": "US-001",
      "title": "Story title",
      "description": "Detailed description",
      "priority": "P1",
      "acceptance_criteria": [
        "Given X, When Y, Then Z"
      ],
      "files": ["src/module.ts", "src/module.test.ts"]
    }
  ]
}

Markdown (prd.md)

## US-001: Story title
**Priority**: P1
**Description**: Detailed description
**Acceptance criteria**:
- Given X, When Y, Then Z
**Files**: src/module.ts, src/module.test.ts

Workflow per story

For each story in priority order (P1 → P2 → P3):

1. LOAD - Load the story

• Read the story from the PRD file
• Display the title and description
• Check prerequisites (files, dependencies)

2. IMPLEMENT - TDD cycle

• Write the tests first (RED)
• Implement the minimal code (GREEN)
• Refactor if necessary (REFACTOR)
• Verify the acceptance criteria

3. COMMIT - Save

• Run the tests: npm test / pytest / go test
• If tests OK: commit with feat(scope): US-XXX description
• If tests KO: STOP and report the blocker

4. REPORT - Update the state

• Mark the story as done in .claude/output/batch/progress.json
• Log the time and modified files
• Move on to the next story

Progress file

Automatic save in .claude/output/batch/progress.json:

{
  "started_at": "2026-03-23T10:00:00Z",
  "stories": {
    "US-001": { "status": "done", "commit": "abc1234", "files": ["..."] },
    "US-002": { "status": "in_progress" },
    "US-003": { "status": "pending" }
  }
}

Resume after interruption

If progress.json exists, resume at the last in_progress or pending story.

Guardrails

• Maximum 10 stories per batch (beyond that, split)
• STOP if 2 consecutive stories fail
• Each story must pass the tests before continuing
• Commit after each story (no giant commits)

---

IMPORTANT: Each story is an atomic commit. Do not accumulate changes.

NEVER continue if the tests fail on a story.

YOU MUST save the progress in .claude/output/batch/progress.json.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: christopherlouet
• Source: christopherlouet/claude-base
• License: MIT
• Homepage: https://christopherlouet.github.io/claude-base/

Install and usage instructions live in the source repository linked above.