Anthropic's Opus 5.5 guide swaps 'think carefully' for finish lines and stop rules

Opus 5.5 thinks before every reply and picks how much thinking the task deserves, according to a dev.to write-up of the guide ^[4]. A "think carefully" line asks the model for something it already does. The write-up does not give a latency figure for Anthropic's chat-product test or say how quality was judged ^[3].

Treat that test as a claim about a chat workload. For it to transfer, a team's tasks would need to resemble chat traffic. The quality check would also need to include the hardest prompts, where a nudge toward more reasoning was most likely to matter. The write-up's author says multi-week experiments in live repos confirmed the guide's core claim, and adds that the numbers and recommendations are the guide's ^[18]. I'd still compare outputs with and without the line on a team's own saved prompts before deleting it everywhere.

The prompt shape is the better engineering in the guide. Osmani's example puts the whole task in one message: "Migrate the payment endpoints from the old client to the new one." ^[6] The finish line follows: "Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes." ^[7] Then one stop condition: "Stop and ask me only if a test fails for a reason you can't explain." ^[8] Each clause of that finish line is something the model can check on its own: a search for the old client, a deleted file, a passing test run. The shape fits the long, lightly supervised repository work where the guide places the model's biggest gains ^[9].

For the mid-task stall, the write-up points to a two-line CLAUDE.md rule from Osmani ^[10]. The first line reads: "When a step doesn't need my input, keep going. Put status notes in the same message as your next action." ^[11] The second sets the exits: "Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository." ^[12] For a single stall, the fix is the lowest-tech one in the guide: reply "continue" ^[13].

A keep-going rule means fewer stops. So the write-up keeps a human checkpoint before anything risky or hard to undo, and leaves permission prompts on for destructive commands ^[14]. I think that split is right. The CLAUDE.md line is an instruction the model interprets. The permission prompt is a setting in the tool, and it still applies when the model misjudges what counts as destructive.

For audits and migrations across a large codebase, Osmani suggests a fan-out: "Give each service to its own subagent. When a subagent reports back, check its evidence before you accept it." ^[15] The run ends in one table listing each service, whether it is affected, and the evidence ^[16].