Run one specific task without a roadmap or full planning cycle. The fastest path to getting something done.
When to use this: You have a specific, well-defined task — fix this bug, add this endpoint, update this config — and you want Claude Code to just do it.
The pattern¶
Write a single task in .dev/sprint.md and run autopilot sprint --auto-approve ..
mkdir -p .dev---
name: my-project
approved: true
max_budget_usd: 2.0
---
### [ ] fix-import-error
The import at `src/app/utils.py:12` fails because `requests` is not listed
in `pyproject.toml` dependencies. Add it and verify:
```bash
uv add requests
uv run python -c "from app.utils import fetch_data; print('ok')"Done: the import succeeds and uv run pytest tests/test_utils.py passes.
Execute:
```bash
autopilot sprint --auto-approve .The --auto-approve flag bypasses the manual approval check since you've already set approved: true in the frontmatter.
Task writing tips¶
The worker agent only knows what you put in the task body. Good tasks include:
- Exact file paths —
src/app/models.py, not "the models file" - Commands to run — so the worker can verify the work
- A
**Done**:criterion — a clear, verifiable statement of what "complete" looks like
### [ ] add-health-endpoint
Add a `GET /health` endpoint to `src/api/routes.py` that returns
`{"status": "ok", "version": "1.0"}` as JSON.
Register it on the FastAPI app in `src/api/app.py`.
**Done**: `curl http://localhost:8000/health` returns `{"status":"ok","version":"1.0"}`.Budget and retries¶
For a single task, a budget of $1–3 is usually plenty. Set max_task_attempts: 1 if you don't want retries:
---
approved: true
max_budget_usd: 2.0
max_task_attempts: 1
---Checking the session¶
Every task spawns a Claude Code session named autopilot/<project>/worker. After a run, you can resume it from Claude Code's session history with /resume to see exactly what happened.
Next steps¶
For more complex work, see Plan & Sprint — let autopilot break the work into tasks for you.