Prompting Claude Opus 5.5: The Old Habits Anthropic’s Own Guide Says to Drop

Anthropic’s own Opus 5.5 guide flags a few prompting habits that cost speed or tokens. Here are the changes to make, with the exact prompts Anthropic suggests.

Reviewed by Mindloop Editorial on . Every link in the Sources section at the end was re-checked that day.

Graphic of an API card labelled opus-5.5 with a crossed-out prompt line, illustrating Anthropic's prompting advice for Claude Opus 5.5.

Here is a small irony from Anthropic’s own documentation: the page that teaches you to prompt Claude Opus 5.5 mostly tells you what to stop writing. Delete the “think carefully” line. Stop dialing thinking up with words. Quit asking for the reasoning to be written out.

The good news is that you do not need to rewrite everything. Anthropic says existing Claude Opus 5 prompts “should perform well without changes.” This guide covers the handful of old habits that, according to Anthropic’s guide, cost you speed or tokens on Opus 5.5, and what to do instead.

What you need

  • Access to Claude Opus 5.5 (API ID claude-opus-5-5), through the API or any app where you control the system prompt. Anthropic lists the price at $4 per million input tokens and $20 per million output tokens.
  • A handful of your own real prompts to test with. Anthropic’s guide repeatedly says to test on your own tasks, and so should you.
  • Context: Opus 5.5 was announced on 22 September 2026. Newer models, covered in our Claude Fable 5.2 explainer, have their own prompting pages. Everything below comes from Anthropic’s Opus 5.5 guide, so don’t assume it transfers.

Anthropic says Opus 5.5 produces output tokens more than 30 percent faster than Opus 5. The habits below are about not giving that back.

Step 1: Delete the “think carefully” line

Many system prompts for chat apps carry a line telling Claude to reason hard before answering. Anthropic’s advice for Opus 5.5 is to consider removing it, because the model decides how much to think and the effort setting is the main control.

The evidence offered: in Anthropic’s testing in a chat product, removing such a line “made replies start sooner, with no clear decline in the quality of the reply.”

Step 2: Turn the thinking dial, not the prompt

If you want less thinking, lower the effort setting first. Anthropic’s guide says lowering effort “reduces thinking, and with it cost and latency, more reliably than prompt instructions do.”

Two related details:

  • Opus 5.5 defaults to medium effort (Opus 5 defaulted to high). Anthropic tells you to set effort explicitly instead of carrying over your old value, since at the same level the new model tends to think more per turn, especially at xhigh and max.
  • Thinking is always on. Sending thinking: {"type": "disabled"} now returns a 400 error. If your old setup relied on that, start at low effort and measure.

Step 3: Tell it to let old answers rest

In long chats, Opus 5.5 sometimes goes back over an earlier answer while thinking about your new message, even a short follow-up. That adds delay. Anthropic suggests ending the system prompt with:

Code
Once you have answered something, treat that answer as done. On later turns, focus your thinking on what the user is asking now, and don't go back over an earlier answer unless the user asks about it or points out a problem with it.

Anthropic says this reduced thinking on follow-up turns without affecting quality in its tests. It also warns the model might become less likely to flag its own earlier mistakes, so skip it for long analyses.

Step 4: Don’t treat “stopped talking” as “finished”

This one bites agent builders. On long jobs, Opus 5.5 posts progress updates, and some of those end the turn with text instead of a tool call. A simple loop reads that as “done” and quits halfway.

Anthropic’s fixes:

  1. Keep the task’s parts in a checklist the model updates, such as a to-do tool or a file.
  2. If a turn ends with items open, send a short nudge, for example: “Your task list still has open items: migrate the remaining two endpoints and update their tests. Continue with them. If one is blocked, say what is blocking it.”
  3. Cap automatic nudges at two or three, so a stuck run ends and a human can look. If you want to build agents that load their own instructions, our Claude Code agent skills guide is a good companion.

Step 5: Say “look around first” for multi-app jobs

Opus 5.5 “tends to get to work quickly.” That is fine until the rule you needed lives in an old email thread or on another spreadsheet tab. Anthropic’s one-sentence fix:

Code
Before taking any action, explore broadly with tool calls: list and open the emails, documents, spreadsheet tabs, and records across the available apps that could be relevant to this task, including ones the task does not explicitly mention, and use what you find.

In Anthropic’s tests this completed noticeably more multi-app tasks correctly, at the cost of slightly more tool calls and tokens. It also warns you to keep untrusted content out of the records it searches.

Step 6: Name the design habits you hate

“Avoid a generic AI look” does not work well. Anthropic says such an instruction “mostly swaps one default for another.” Naming specific patterns does. The example prompt in Anthropic’s guide includes this sentence:

Code
Do not use a cream or off-white background, italic accent words in headlines, numbered "01/02/03" section labels, monospace labels, or pill-shaped buttons.

Check which styles slipped through, then extend the list.

Old habit versus new habit

Old habitWhat Anthropic’s guide says
“Think carefully” in chat promptsConsider removing it
Reusing your Opus 5 effort valueSet effort explicitly, start at medium
Asking for less thinking in wordsLower effort first
Any text-only turn means doneCheck a task list before stopping
Vague “make it look good”List the exact patterns to avoid

Common errors and fixes

  • HTTP 400 on thinking. Opus 5.5 rejects thinking: {"type": "disabled"} and manual budget_tokens. Omit the field or use adaptive thinking, and control depth with effort.
  • Replies cut off. Thinking counts toward max_tokens. Anthropic reports that 128,000, the model’s maximum, worked well for long agentic coding turns.
  • A silent app. Progress notes now arrive as thinking blocks with empty text by default. Set display: "updates" (a beta feature) to receive summaries.
  • stop_reason: "refusal". Opus 5.5 runs safety classifiers. A prompt that asks it to write out its reasoning can be declined with a reasoning_extraction category, so remove that instruction and read the summarized thinking blocks instead.

The takeaway

Anthropic’s own testing found Opus 5.5 at default medium effort matched or beat Opus 5 at high on coding tasks, in fewer steps. The cheapest upgrade may be deleting lines. As of 15 November 2026, these notes reflect Anthropic’s guide, which can change, and none of the prompts above were run for this article.

Sources

Links last checked on 15 November 2026.

Written by Mindloop Editorial

The Mindloop Experience editorial team writes clear, carefully sourced explainers and how-tos on AI and technology, software and the business of tech, for anyone who likes to know how things actually work.

Meet the editorial team

Responses

Write a response

More from Mindloop Experience

See all from Mindloop Experience

You can clap up to 50 times.