Skills in your repo

What a job does is written as a skill: a Markdown file of instructions in your repo. The job itself only names it.

Where skills live

• Directoryyour-repo/

  • Directory.claude/

    • jobs.json the jobs record: when each job runs
    • Directoryskills/

      • Directoryopen-issues-digest/

        • SKILL.md what the digest job does
      • Directoryweekly-commits-summary/

        • SKILL.md
    • Directorymemory/ optional: Claude’s shared memory

      • …
  • …your own files

Everything the template adds sits under .claude/, so the rest of your repo stays as it was.

What a skill looks like

Claude writes the skill when you ask for a job. A skill is plain steps, for example:

.claude/skills/open-issues-digest/SKILL.md

---
name: open-issues-digest
description: List the repo's open issues with a one-line summary each.
---

# Open issues digest

1. List the open issues in this repo with `gh issue list --state open`.
   Pull requests are not issues; ignore them.
2. Write one line per issue, oldest first:
   `#<number> <title>: <one-line summary of what is asked or reported> (<url>)`.
   Base the summary on the issue's body, not just its title.
3. Start with a heading giving today's date and the count.
4. Mark the task done with the digest as its summary, or blocked with the reason.

The last step matters: a job’s run must close its task. See No overlap, no backlog.

How a run uses it

1. The job’s prompt names the skill: “Run the open-issues-digest skill.”
2. The Runner pulls the repo, so it has what is pushed.
3. It runs that skill from the repo’s .claude/skills/, and does what it says.

Since the Runner pulls first, a change to a skill takes effect at the next run once it is pushed. A change you haven’t pushed doesn’t.

Change a skill

• Ask Claude in CloudCLI: “the digest should also list pull requests”. Claude edits the skill, commits, pushes and tests the job.
• Edit it yourself, anywhere: on GitHub, in your editor, in CloudCLI. Commit and push.

Every change is a commit, so you can review it, and revert it.

Your repo's own settings apply

A job runs in your repo, so the repo’s own Claude settings apply to every run, as in the chat: its CLAUDE.md, .claude/settings.json, hooks and tool servers.

Skills you can reuse

A skill is not tied to this template. It is a Claude Code skill: Claude Code finds it in .claude/skills/ wherever you open the repo, and you can run it yourself by asking for it. Codex reads the same file when it runs a job.

In a product repo or a separate one

• In a product repo, the skills and the record sit in its history, next to the code they work on.
• In a repo made for jobs, they stay apart from your code. Good for jobs that aren’t about code: reports, checks, drafts.