Claude Opus 5.5 prompting guide: effort, agents, refusals and templates

To prompt Claude Opus 5.5 well, keep the prompts you used on Claude Opus 5, set the effort level before you touch the wording, and then add one targeted line to the system prompt for each behavior you want to change. Anthropic's guide says existing Claude Opus 5 prompts "should perform well without changes", and that effort is "the first setting to adjust when trading off intelligence, latency, and cost". Most of the new advice is about agents: keeping an unattended run going, getting progress updates out of long turns, and pacing a team of agents. Everything below comes from Anthropic's "Prompting Claude Opus 5.5" guide and the model's API reference, checked on October 3, 2026.

Request rules shared with Claude Fable 5.1, such as no assistant prefill and no custom temperature, are covered in the Claude Fable 5.1 prompting guide. This article covers what is specific to Claude Opus 5.5.

Do Claude Opus 5 prompts work on Claude Opus 5.5?

Yes. Anthropic says they "should perform well without changes", and its Claude Opus 5 prompting patterns "remain a reasonable starting point". What changes is how the model spends its effort. Claude Opus 5.5 writes output tokens more than 30 percent faster than Claude Opus 5 and tends to finish the same task with fewer tokens.

Anthropic names four areas where the model improved, and they decide which parts of this guide matter to you:

• Agentic coding and code review. Multistep work in a real repository, long autonomous runs with subagents, and code review that catches more bugs with fewer false alarms.
• Knowledge work. It is much less likely to state a wrong figure or cite the wrong source, and it catches small inconsistencies in large inputs, such as a date that falls on the wrong weekday.
• Communication. Progress updates and final reports say plainly what it did, what it found and what it needs from you.
• Charts, diagrams, screenshots and computer use. It reads visual material more accurately than Claude Opus 5, even at its lowest effort.

If your prompts are mostly chat, read the effort and chat sections. If you run agents, the agent sections are where the new advice is.

What effort level should I use for Claude Opus 5.5?

Start at medium, which is the default on Claude Opus 5.5 (Claude Opus 5 defaults to high). Set it explicitly and test several levels against your own tasks instead of carrying over the value you used on Claude Opus 5. Effort names do not map to the same amount of thinking across models: in Anthropic's testing, Claude Opus 5.5 at medium "matches or exceeds Claude Opus 5 at high on coding and knowledge-work evaluations", and on several coding evaluations low comes close to that at a much lower cost.

Setting medium explicitly behaves exactly like leaving effort out.

At the same level, Claude Opus 5.5 tends to think more per turn than Claude Opus 5, especially at xhigh and max. Keep three things in mind:

• Leave room in max_tokens. Thinking counts toward max_tokens even when you do not see it, so a limit sized for Claude Opus 5 with thinking off can cut replies short. For long agentic coding turns, Anthropic found a max_tokens of 128,000, the model's maximum, worked well.
• Lower effort before you write "think less". Lowering effort reduces thinking, cost and latency more reliably than prompt instructions do.
• Keep effort fixed if you rely on prompt caching. Changing the top-level effort value between requests invalidates the prompt cache.

Effort is set in output_config:

{
  "model": "claude-opus-5-5",
  "max_tokens": 128000,
  "output_config": {"effort": "medium"},
  "messages": [{"role": "user", "content": "Review this pull request for bugs and explain each one."}]
}

Can I turn off thinking on Claude Opus 5.5?

No. Claude Opus 5 accepted "thinking": {"type": "disabled"} at high effort or below; Claude Opus 5.5 does not, and the request returns a 400 error. Thinking is adaptive and always on. If your integration ran with thinking disabled, Anthropic recommends four changes:

1. Start at low effort and measure. At low the model keeps its thinking short. If time to first token still matters, a system prompt line such as "Answer directly without deliberating." can cut thinking further, but check quality when you add it.
2. Remove instructions that stood in for thinking. If your prompt asked the model to write out its reasoning in the reply, remove that line and read the reasoning from summarized thinking instead. A prompt that pushes the model to reproduce its reasoning in the reply can now be declined (see the refusals section below).
3. Re-test old workarounds. Rules you added to avoid artifacts of running without thinking may no longer be needed. Remove any rule that tells the model not to think.
4. Read the response by block type. A response may or may not start with a thinking block, so check each block's type instead of assuming the first one is text.

To see the reasoning, ask for summarized thinking:

{
  "model": "claude-opus-5-5",
  "max_tokens": 16000,
  "thinking": {"type": "adaptive", "display": "summarized"},
  "output_config": {"effort": "low"},
  "messages": [{"role": "user", "content": "Which of these three invoices has a total that does not match its line items?"}]
}

How do I keep an unattended Claude Opus 5.5 agent from stopping early?

On long tasks, Claude Opus 5.5 keeps the user informed as it works, and some of those updates end the turn with text instead of a tool call (stop_reason: "end_turn"). An agent loop that treats every such turn as "task finished" stops halfway. Anthropic's fix has a harness part and a prompt part.

In the harness, treat a text-only end of turn as a report, not as proof the work is done. Keep the task's parts in a checklist the model updates, such as a to-do tool or a file. If a turn ends with items still open and no blocker named, send a short user message that lists them:

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.

Stop after two or three automatic continuations on the same task, so a run that is genuinely stuck ends and can be reviewed. If a background command or subagent is still running, wait for it and return its output as the next user message.

In the system prompt, name the kinds of early stop you do not want. Anthropic's example, written for agents that run fully unattended, goes at the end of the system prompt from the first request of the session. Adding it partway through changes the system prompt and invalidates the conversation's earlier thinking blocks.

A standing instruction from the user, the person you are working for. It is about how your turns end. A message with no tool call in it ends your turn, and the work stops there until you are asked to continue. The user has seen you end turns in four ways while work they asked for was still owed, and does not want any of them. One: a long summary of what was done that closes by announcing the next step and has no tool call, so the next thing never starts. Two: an offer to carry on with something unless the user would prefer otherwise, which stops to wait for an answer the user was not going to give. Three: a list of decisions for the user when, by your own account, none of them blocks the rest of the work. Four: deciding that this is a good place to report, because the turn has been long or a milestone is done. Status notes are welcome, and so are your recommendations on open decisions, but put them in the same message as your next tool call and carry on with whatever does not depend on the user's answer. If you notice yourself inviting the user to redirect you or offering to wait, delete it and do the next thing. The stops the user does want are the ones where nothing can move without them, or where the thing blocking you is deliberately protected from you. This does not override the need for confirmation on risky or destructive actions.

Leave this addition out of human-in-the-loop applications, where someone is there to answer, and keep your own confirmation step for risky or irreversible actions. Expect somewhat more tool calls and output tokens per task.

How do I get progress updates from long Claude Opus 5.5 agent runs?

Between tool calls, Claude Opus 5.5 writes short progress notes: what it just found and what it will do next. On this model those notes come back as thinking blocks, not text blocks, and their text is empty by default, so a client that shows only text blocks can look silent during a long turn. Anthropic documents a beta display mode that returns a summary of each note; the Claude Opus 5.5 API reference lists the thinking.display values the endpoint documents.

Three prompt and harness levers work without any special setting:

• Give the model a tool for verbatim content. If it may need to hand the user something exact mid-turn, such as a code snippet, give it a simple "send a message to the user" tool and say to reserve it for that. Declare the tool from the first request; adding it later invalidates earlier thinking blocks.
• Ask for updates at fixed points. For example, a one-line statement of intent before the first tool call and a short recap at the end. The model follows such instructions closely, which helps most when a person is watching.
• Nudge a quiet turn from the harness. Count tool-calling steps in a row that give the user nothing to read. After several, five for example, append a reminder after the latest tool results, and stop after two or three reminders. Anthropic's wording:

The user hasn't heard from you in a while — say in a few words what you're doing, then continue.

In Anthropic's testing on agentic coding tasks, this reminder roughly halved the share of tasks with a long silent stretch, with no measurable change in cost.

How do I make a Claude Opus 5.5 agent check context before acting?

Claude Opus 5.5 tends to start work quickly. In automations that span email, documents, spreadsheets and CRM records, the fact a task depends on often sits somewhere the request does not mention, such as a policy in an old email thread or a rule on another spreadsheet tab. One system prompt sentence makes the model look around first:

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 testing on multi-app tasks, the model completed noticeably more of them correctly with this line, at both medium and max effort, for slightly more tool calls. Because the line tells the model to act on what it finds, keep untrusted content out of the records it searches.

How do I make a team of Claude Opus 5.5 agents finish sooner?

Claude Opus 5.5 pays close attention to elapsed time. In a setup where a lead agent hands work to subagents, give it a time budget: have your harness add a short line to the end of each message it sends back to the model, such as elapsed 340s / 1200s. The model paces its work to finish inside the budget and usually finishes well before it, so set the budget somewhat above the time you want spent. If you cannot estimate a budget, show the elapsed time alone and add this sentence to the system prompt:

Time matters here: do not spend time that can be avoided, and the earlier a correct result is obtained, the better.

The budget is advisory and nothing stops the model at the limit, so keep your own timeout if you need a hard stop. A tighter budget is not the same as lower effort: lower effort reduces the work itself, while a budget mostly keeps more agents working in parallel. Under time pressure the model may search and verify a little less, so check answer quality on your own tasks.

Should a chat system prompt tell Claude Opus 5.5 to think carefully?

No. Anthropic suggests removing "think carefully before answering" lines from chat system prompts: the model decides how much to think, and effort is the control. In Anthropic's testing in a chat product, removing such a line made replies start sooner with no clear drop in quality.

Two more chat lines are worth knowing. One tells the model to treat earlier answers as settled, so follow-up turns start faster. The other marks text the user pasted in, so the model does not follow instructions hidden inside an email or web page. Both lines, word for word, are in the Claude Opus 5.5 section of the Fable 5.1 prompting guide. Leave the "settled answers" line out of long analyses and agentic tasks, where a later step can reveal a mistake in an earlier one.

How do I get better answers about charts, diagrams and screenshots?

Re-test first. Claude Opus 5.5 reads dense charts, flowcharts and calendar screenshots far more accurately than Claude Opus 5 without extra tooling; in Anthropic's testing, even at its lowest effort it read values off dense charts more accurately than Claude Opus 5 did at its highest. Scaffolding you built for older models may no longer be needed.

For the densest inputs, two things still help:

• Higher-resolution images, most of all for technical drawings.
• Image tools. Let the model run as an agent with a container that holds the raw images and has PIL or OpenCV, so it can crop, zoom and measure. If that is too much, a cropping tool alone still helps. The model uses these tools better at higher effort.

Without tools, raising effort improves its reading of technical drawings but does little for charts. Images go in a user turn as image blocks with a URL source; the API reference shows the format.

How do I stop Claude Opus 5.5 frontend output from looking generic?

Name the patterns you do not want. Without design direction, Claude Opus 5.5 falls back on a few default styles, and a general line such as "avoid a generic AI look" mostly swaps one default for another. Anthropic's example lists concrete things to avoid:

Output a vanilla HTML/CSS personal website with placeholder data. 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.

Work iteratively: look at which styles the first result used instead, and add them to the list.

Why does Claude Opus 5.5 return stop_reason "refusal"?

A safety classifier declined the request. Claude Opus 5.5 runs classifiers for biology, cybersecurity and reasoning extraction, and a decline arrives as a normal response with stop_reason: "refusal" and a stop_details object that names the category.

For refusals of harmless requests in general, see why Claude refuses a harmless request in the Fable 5.1 guide.

Frequently asked questions

Is there an official Claude Opus 5.5 prompting guide?

Yes. Anthropic publishes "Prompting Claude Opus 5.5" in its developer documentation, alongside a general prompting best-practices page that applies to all current Claude models. This article summarizes the Opus 5.5 guide as of October 3, 2026, with the templates quoted word for word.

What is the default effort on Claude Opus 5.5?

medium. Claude Opus 5 defaults to high. The levels are low, medium, high, xhigh and max, set in output_config.effort.

How many output tokens can Claude Opus 5.5 write?

Up to 128,000 tokens per request, and that budget includes thinking. The context window is 1M tokens.

Can I prefill Claude Opus 5.5's answer or set the temperature?

No. Assistant prefill and any temperature other than the default return a 400 error, and forcing a specific tool is not supported. Describe the format you want in the prompt instead. The Fable 5.1 prompting guide explains what to use in place of each.

Is Claude Opus 5.5 or Claude Fable 5.1 better for agents?

It depends on the task and the budget; the Claude model comparison covers which to choose. The prompting advice differs mainly in defaults: Claude Opus 5.5 starts at medium effort, Claude Fable 5.1 at high.

Try the prompts

Run these templates on Claude Opus 5.5 with a SeedRouter key, through the Anthropic SDK or the OpenAI Chat Completions and Responses formats. The Claude Opus 5.5 API reference lists every parameter, and the model page shows current per-token prices. SeedRouter is pay as you go: top up once, credits never expire, and failed requests are not charged.