Opus 5.5: delete “think step by step” and name the finish line

What happened

Anthropic published a guide to getting the most out of Opus 5.5 in the Claude apps and in Claude Code. It reached the front page of Hacker News on October 3. The guide's starting point is that Opus 5.5 keeps working on its own for longer than earlier Opus models, thinks before every reply, and reports more plainly what it did. So a prompt should describe the finished result and when to stop, not every step on the way.

The five changes it recommends:

• Delete "think carefully" and "think step by step" from your prompts and saved instructions. Opus 5.5 always thinks before it answers and decides how much on its own. Anthropic says that in its testing, removing such a line made replies start sooner with no clear drop in quality. If you want a quick answer, say "answer directly". In Claude Code, change the effort setting instead.
• Give the whole task in one message, with a finish line. Say what "done" means, such as "the tests pass", and name the one case where it should stop and ask you.
• List the design habits you don't want. Without direction, Opus 5.5 falls back on a few default styles, and "avoid a generic look" only swaps one default for another. A list of specific patterns to leave out works better.
• Keep the task list in a file on long runs. When the context window fills, Claude Code summarises older turns. A checklist in a file survives that, and you can read it to see where the run is.
• Ask it to mark what it couldn't confirm in research and analysis, and to say where it looked.

What it means if you use Claude every day

Most saved prompts, project instructions and CLAUDE.md files written over the last two years carry lines like "think step by step" and long numbered procedures. They were useful on older models. On Opus 5.5, Anthropic's own testing says the thinking line only delays the answer. Searching your Claude projects and CLAUDE.md files for those lines is a five-minute job.

The bigger change is how you hand over work. Writing a precise finish line now matters more than writing precise steps: "every endpoint uses the new client and the test suite passes" tells the model where to stop. A list of steps tells it how to start and leaves the end to guesswork. The same goes for reviews and research: ask for blocking problems only, and for an explicit list of what it could not check.

Try it

Task: [the whole job in one sentence]
Done when: [for example, every file is migrated and the test suite passes]
Stop and ask me only if: [one condition]
Keep a checklist in TASKS.md and tick items as you finish them.

Build [a landing page for my bookkeeping service].
Do not use: [cream or off-white backgrounds, italic accent words in headings, numbered section labels, pill-shaped buttons, monospace labels].
After the page, list every design choice you made in one line each.

[Your research question]
Mark anything you couldn't confirm, and say where you looked for it.

Full versions with more structure are in the library: Long Run Task Brief and Website Brief With a Ban List. More in Claude Code prompts.

Sources

• Anthropic, Getting the most out of Opus 5.5 in Claude and Claude Code
• Hacker News, discussion thread, October 3, 2026