Ask Claude for a job

You create jobs by asking Claude in CloudCLI, in plain words. You don’t need to name a skill, write cron or open Paperclip.

Where to ask

Open a chat in the repo the job is for: in CloudCLI’s sidebar, pick the repo’s project, then New Session. A job belongs to the repo of the chat you create it in.

If you ask from somewhere else, such as the home folder or a repo that isn’t connected, Claude asks which repo the job is for.

What to say

Say what should happen and when. Add anything that isn’t obvious:

• What: “list the open issues with a one-line summary each”, “check that our site answers”, “draft release notes from last week’s merged pull requests”.
• When: “every weekday at 8:00”, “every Monday at 9:00”, “on the first of each month”.
• Time zone, only if it isn’t yours: “at 9:00 Paris time”.
• Where the result goes, if not only in Paperclip: “and post it to our Slack channel”. See Outside services.

Some requests that work well:

Every weekday at 8:00, list the open issues in my repo with a one-line summary each.
Every Monday at 9:00, summarize last week's commits in my repo.
Every morning at 7:00, check that https://example.com answers, and mark the task blocked if it doesn't.
On the 1st of each month at 7:00, check the repo's dependencies for new major versions.

Not sure what to ask? Say hi. Claude looks at your repo and suggests two or three jobs that fit it.

What Claude does

1. Looks at your repo and the jobs you already have.
2. Writes the skill: the job’s instructions, as .claude/skills/<name>/SKILL.md in the repo. It commits and pushes it, since the Runner pulls before every run.
3. Creates the job in Paperclip: a title, a schedule in your time zone, and the prompt “Run the <name> skill.”
4. Records the job in the repo’s .claude/jobs.json, then commits and pushes that file.
5. Tests it: runs the job once, waits for it to finish and reads its log.
6. Reports the result, with links to the job, the task and the run in Paperclip.

You see each step in the chat. Read the skill Claude wrote: it is what every run will do.

Good to know

• One job, one skill. The prompt only names the skill, plus at most one line of arguments. To change what a job does, change its skill.
• The result stays in Paperclip by default, on the run’s task. Claude can also send it somewhere, such as a chat or an inbox, if you ask. See Outside services.
• A job that needs a key (an API key, a webhook address) gets it as a secret on the Runner, which you add in Paperclip. Claude never asks for the value in the chat. See Secrets.
• Codex: to run a job with OpenAI’s Codex instead of Claude, say so: “run it with Codex”. See Codex.

What Claude may do without asking

So that a first job doesn’t stop at several permission prompts, the template approves a few actions once, in Claude Code’s user settings on CloudCLI:

• the jobs script;
• editing and writing files, and waiting on a command;
• basic git: add, commit, push, pull, status, diff, log;
• reading issues and pull requests with gh.

Anything else still asks. A rule you remove later stays removed.