Execute Story

Using agents

Work on your own by default. No subagents, background agents or Workflow runs without asking — ask briefly ("May I use agents for this? Reason: …") and wait for an explicit yes.

Exception: if the user has already cleared agents for this story, or asked for them themselves, that clearance holds for the story and you do not ask again.

Setup

If this is the first story: check whether we need a branch — we do not want to work on master.

Load the context

Read ALL the stories passed in, thoroughly and completely (lessons learned, handover, ticked-off tasks) so you hold the full context:

$ARGUMENTS

Implementation

Identify the next open story (the first one with unfinished tasks) and implement ONLY THAT ONE:

1. Analyse what is needed and start once your confidence is sufficient
2. Update the tasks in the story document as you go (tick the checkboxes)
3. Fill in at the end:
   • Lessons Learned (what went well, architecture decisions, technical debt)
   • Handover to the follow-up story (context, relevant files, starting prompt)

Closing out (IMPORTANT — after EVERY story)

Once the story is done, these steps in order:

1. Run the tests

Run the project's relevant tests and make sure nothing broke. Which ones those are is in the project's CLAUDE.md or its package.json scripts — when in doubt, the full unit suite plus the E2E/integration tests covering the flows you touched. If a lane needs running infrastructure (database, dev server), check that first instead of starting blind.

2. Auto-commit

Create a commit whose message references the story (e.g. feat(feature): implement Story N — short description).

3. Completion report

Give a compact closing report:

• Story: name and number
• What was done: 3-5 bullet points
• Key files: new and changed files
• Next story: name + short description (if there is one)
• Notes: open points, workarounds, anything unusual

4. Testing notes

Say clearly what is already checked and what still needs hands:

• Tested by me: which automated tests ran, which functional checks you did yourself
• Recommended to check manually: what the user should look at (UI interactions, edge cases, integration points)
• Could still test: what else is feasible, if they want it

5. STOP

End the work here. Do NOT implement the next story automatically. The user decides how to continue (new session, next story, a break, a review).