Claude Operator: Prompt to Autonomy · 20 min · 140 XP
Isolation, scheduling, CI, and graduating
Where an autonomous run should execute, and how to earn each step of trust.
Isolate the work. Anything that changes code runs in a temporary branch or worktree, never on the main branch. An isolated diff is one you can discard entirely, which is what makes reviewing it a real decision rather than a formality.
Schedule something harmless first. A read-only report on a timer, with a documented and tested way to disable it. Scheduling is where a workflow stops being something you run and becomes something that runs, so the first one should be something you'd be relaxed about running wrong at 3am.
In CI, the environment is the control surface. Pin versions so a run is reproducible. Give the job the minimum token permissions it needs. Bound the commands. And treat anything from a pull request (title, description, branch name, diff) as untrusted input, because on a public repository it's written by a stranger who knows your automation reads it. That's the same prompt injection from Level 3, in a place where nobody is watching.
Then graduate in four stages. Manual. Suggested: it proposes, you do. Approved execution: it acts after you approve each action. Bounded autonomy: it acts within limits and reports. Move up when the current stage has been boring for a while, and define in advance what sends you back down: a wrong action that reached production, a budget surprise, an alert nobody understood.
Measure quality across the stages. Keep a small evaluation set (a dozen real inputs with known-good outcomes) and re-run it after every prompt or skill change. Without it, a change that fixes today's case and breaks three old ones looks exactly like an improvement, and you won't find out until the workflow is autonomous.
Practice. Run a code-changing practice task in a temporary branch or worktree and confirm the main branch is untouched. Schedule a read-only local report and test the disable path. Design a CI job with a pinned environment, minimum token permissions and bounded commands, then feed it malicious task text and confirm your boundaries outrank the content. Build a small evaluation set and use it to catch at least one regression after a prompt change. Then write your four-stage rollout with explicit rollback criteria.
Loading your workspace…