Opus 5.5 guide tells users to name the finish line and write stop rules into CLAUDE.md

The guide's central advice for long runs is to give the whole task in one message and say what "done" looks like. Its example prompt ends with a stop condition: "Stop and ask me only if a test fails for a reason you can't explain." It also reports that Opus 5.5 sometimes stops mid-task to summarize or offer to continue, and suggests a CLAUDE.md rule: "When a step doesn't need my input, keep going." The same rule says to stop and ask only before anything destructive, and the guide says to keep permission prompts on for destructive commands.

Other specifics: delete "think carefully" lines, because the guide says the model already thinks before every reply and that removing such a line in a chat product made replies start sooner "with no clear drop in quality" (its own testing, no figures given). For large audits, split work across subagents and have the model check each one's evidence before accepting it. Keep the task list in a file such as TASKS.md so it survives context summarization. The guide also says requests to reproduce internal reasoning in the reply can be declined as a flag category. Its claims about long unattended runs and bug-catching rest on unnamed early testers, and no measurements are given.

Why it matters: If you run Opus 5.5 unattended, the stop-and-continue rule, the finish-line prompt and the file-based task list are changes you can make to a harness this week. Treat the performance claims as the vendor's account of its own model; the guide offers no numbers to check them against.