Opus 5.5 in Claude Code: the 4 prompt changes that let it work longer, and the command that fixes your old instructions

With Opus 5.5, Claude Code can work on its own for much longer than before. To get that, though, you have to change your prompts: drop "think carefully", give a precise goal, say what you don’t want and have it keep a checklist in a file. For the instructions and skills you have already written, the /claude-api prompt-audit command finds the lines meant for older models.

Both come from Anthropic. The first is the official guide Getting the most out of Opus 5.5 in Claude and Claude Code, published on the claude.dev blog on September 22, 2026, the day the model was released. The second is a subcommand of the official claude-api skill, which already ships with Claude Code.

What changes with Opus 5.5 in Claude Code?

Opus 5.5 sustains much longer autonomous sessions than earlier models. According to Anthropic’s documentation, it carries out multi-hour audits and migrations of large codebases, even with parallel subagents and little oversight.

A prompt that worked fine until yesterday now leaves much of that autonomy on the table. The four changes below are the ones I apply when I rewrite a prompt for Opus 5.5.

What are the 4 prompt changes?

1. Why drop "think carefully"?

Opus 5.5 already reasons on its own: thinking is always on. The phrase "think carefully" no longer does anything useful.

Anthropic’s documentation says so explicitly: the main control over how much the model thinks is the effort level, and instructions like "think carefully before answering" can be removed. In Anthropic’s testing on a chat product, removing them made replies start sooner with no clear drop in quality.

2. Why give a precise goal?

With a precise goal, Opus 5.5 knows when to stop and checks on its own that everything works. A vague objective leaves the model to decide when the job is done.

A useful goal is verifiable. For a booking page, I would write: "Done means: I pick a day and time, a confirmation email arrives, and two people can’t book the same slot. Try it yourself before telling me it works." The guide also suggests stating the completion condition at the start of the task.

3. Why say what you don’t want?

If you only ask it to "do a good job", Opus 5.5 always gives you the same style. The model falls back on its defaults.

Anthropic’s guide shows this with frontend work: a generic instruction like "avoid a generic AI look" just swaps one default for another. Listing the specific patterns to avoid works better, and so does extending the list after you see the first result.

4. Why keep a checklist in a file?

Opus 5.5 works on its own for much longer, and as the conversation grows it loses older details. A checklist file stays put.

The checklist does two jobs:

• for you: you always know where the work stands, even hours later;
• for the model: it reasons better about the tasks, because what is done and what is left is always in front of it.

Anthropic also recommends keeping the parts of a task in a checklist the model updates, in a to-do tool or a file. On long tasks in large projects, I add a line at the end of the prompt that asks for exactly this: "Keep a checklist in TASKS.md. Tick off each item when it’s done and add whatever you discover along the way."

What does the /claude-api prompt-audit command do?

If you use Claude Code and wrote instructions or skills for older models, run /claude-api prompt-audit. The command reads through your prompts, skills, CLAUDE.md, commands and subagents, and looks for lines written for earlier models.

Among the patterns it flags:

• all-caps "always" and "never", such as MUST, NEVER, ALWAYS, CRITICAL and IMPORTANT repeated in the same prompt;
• step-by-step procedures ("STEP 1… STEP 2…") on tasks that call for judgment;
• instructions that steer reasoning, such as "think step by step".

For each line it tells you where it is (file and line), why it should go today and how to fix it, with a proposed edit. And it doesn’t touch anything until you ask: the skill’s documentation says edits are applied only on explicit request.

The principle is the same as periodically cleaning up your CLAUDE.md: instructions written for a model age along with it. I covered this too when looking at how Claude 5 changed the rules of context engineering.

Why is it worth rewriting your prompts?

Claude Code stays the same, but with Opus 5.5 the right prompt is different. A prompt built for older models pushes Opus 5.5 to reason where it doesn’t need to, to stop without a clear criterion and to repeat the same style.

Four changes and one command are enough to get started. If your CLAUDE.md is a few months old, now is a good time to audit it.