You have a prompt meant to update two hundred files and you have never run it. Once started it runs autonomously, so it cannot stop to ask you anything. How should you begin?

The principle — an untested prompt at scale is not one chance to be wrong, it is two hundred chances to be wrong in the same way.

Why the key is correct — Anthropic's guidance is to refine your prompt based on what goes wrong with the first two or three files, then run on the full set. The autonomy stated in the question is what makes that binding rather than merely tidy: the task runs on its own and so cannot ask clarifying questions. Anything left ambiguous will be decided by the model, silently, every time. A short first run is the only place that ambiguity surfaces while it is still cheap to fix, and the fix lands in the prompt where it protects the remaining files.

Why the others are wrong — running the whole set and repairing the output afterwards reverses the documented order, and spreads the repair across two hundred files instead of concentrating it in one prompt. Expecting the run to raise its hand mid-way contradicts what autonomous means here. And a first outing is not the place for the hardest case: the advice for unfamiliar delegation is to begin with work that has clear boundaries, which also keeps the diagnosis clean.

Remember this — try it small, fix the prompt, then let it loose.

Sources — Anthropic's Claude Code best-practices, common-workflows and agent-teams documentation.