Anthropic released a guide to Opus 5.5: don’t micromanage the model

Anthropic has released a practical guide to working with Opus 5.5, the new model in the Claude line. The playbook "Getting the most out of Opus 5.5 in Claude and Claude Code" came out on September 22; its author is Eddie Osmani; sources and a photo caption will be added separately.

The main piece of advice is not to micromanage the model. It's enough to describe the task in full in one message and name the finish line right away: "the tests pass," "every endpoint is migrated." After that Opus 5.5 works on its own, and there's no need to break the job into small steps for it. The model can sustain multi-part tasks longer than the previous version: early testers let it write code for hours with almost no supervision.

The guide recommends removing instructions like "think carefully" and "reason step by step" from prompts: Opus 5.5 already thinks before every answer and decides for itself how much to think. In the guide authors' tests, removing such a line sped up the start of responses without any noticeable loss of quality.

A separate section covers managing a long run in Claude Code. The advice is to add a rule to the CLAUDE.md file: keep working until a human's answer is needed, and stop only before irreversible actions — deleting data, force-pushing, or making changes outside the repository. Permissions for dangerous commands, meanwhile, are best left enabled.

The task list is best kept in a separate file like TASKS.md: a long run fills up the context, which later gets compacted, while the file survives it. Large audits and migrations are handed off to subagents, with each one's evidence checked before its result is accepted. When a run ends, you first read what the model expects from you, and only then the report.

Opus 5.5 is the first Opus model with Fable-level biosecurity and cybersecurity: messages that touch these mechanisms are routed to an older model, and the work continues there. To get back to Opus 5.5, you can use model selection, the /model command in Claude Code, or turn off auto-switching in the settings. The guide includes ready-made prompts and a checklist for before a long task.

Translation editor: Салтанов А.