Prompting Claude Opus 5.5

This guide covers the triggering patterns particular to Claude Opus 5.5. For the design’s abilities and API modifications, see What’s brand-new in Claude Opus 5.5. For strategies that use throughout all present Claude designs, see Prompting finest practices.

Claude Opus 5.5 creates output tokens more than 30 percent quicker than Claude Opus 5 and tends to complete the very same job with less tokens. Existing Claude Opus 5 triggers need to carry out well without modifications, and the patterns in Prompting Claude Opus 5 stay an affordable beginning point. Start with the area that matches what you observe:

• Uncertain which effort level to run, or turns run longer and cost more than they did on Claude Opus 5: Calibrate effort
• Your Claude Opus 5 combination kept up believing handicapped: Prompts composed for believing handicapped
• An ignored representative stops partway through a long job after reporting development: Unattended agentic runs
• Demands return stop_reason: "refusal": Safeguard rejections
• Long agentic turns look quiet, or you desire updates at foreseeable points: User-facing development updates
• A representative that works throughout numerous linked apps misses out on details the job didn’t indicate: Explore context in multi-app workflows
• You run a group of representatives and desire it to end up faster: Time signals for multiagent harnesses
• Replies in a chat application start gradually since the design believes at length initially: Thinking guidelines in chat system triggers
• The design follows directions that got here inside text a user pasted: Mark pasted text in user messages
• Responses about thick charts, diagrams, or screenshots miss out on information: Tools for intricate visual inputs
• Frontend output looks generic: Frontend style defaults

Abilities appropriate to triggering

The abilities that matter most for triggering are:

• Agentic coding and code evaluation: The design is greatest on multistep operate in a genuine repository, such as bring a modification through a big code base till its tests pass. In Anthropic’s screening, at its default medium effort the design matched or beat Claude Opus 5 at high effort on such jobs, in less actions and with less tokens. It likewise sustains long-running self-governing work much better than Claude Opus 5, such as multi-hour audits and migrations of big code bases run end to end with parallel subagents and little oversight. Early testers likewise reported more powerful code evaluation, with more bugs captured than on Claude Opus 5 and less incorrect alarms, and it discusses its modifications in plain language.
• Understanding work: The design is much less most likely to mention an inaccurate figure or mention the incorrect source. It’s much better at monetary modeling jobs, such as constructing a monetary design and one-page summary for a deal or finding and repairing mistakes in an assessment workbook, and it captures information that are simple to miss out on in big inputs, such as a date in a long preparation thread that falls on the incorrect weekday or a chart in a slide deck that does not match the hidden figures. The spreadsheets, slides, and files it produces need less modifying before you share them.
• Interaction: Its reports on agentic work, both the updates while it works and the summary when it completes, state clearly what it did, what it discovered, and what it requires from you. See User-facing development updates.
• Charts, diagrams, screenshots, and computer system usage: The design checks out visual product more properly than Claude Opus 5 without additional tooling: in Anthropic’s screening, even at its most affordable effort setting it checked out worths off thick charts more properly than Claude Opus 5 did at its greatest, utilizing a little portion of the output tokens. It is much better, too, where significance depends upon position instead of text: which boxes an arrow links in a flowchart, what altered in between 2 variations of a diagram, or precisely when a conference begins and ends in a calendar screenshot. It’s likewise more trusted at computer system usage, where it runs applications from screenshots over lots of actions: at its default effort it matched the success rate that Claude Opus 5 reached just at a much greater effort setting. See Tools for intricate visual inputs.

Adjust effort

Effort is the primary control for just how much Claude Opus 5.5 believes, and since thinking is constantly on, it’s the very first setting to change when compromising intelligence, latency, and expense. Start at mediumthe default on Claude Opus 5.5 (Claude Opus 5 defaults to highset it clearly, and test a number of levels versus your own evals instead of rollovering the setting you utilized on Claude Opus 5. Effort level names do not represent the very same quantity of believing throughout designs: in Anthropic’s screening, Claude Opus 5.5 at medium matches or goes beyond Claude Opus 5 at high on coding and knowledge-work assessments, and on numerous coding examinations low comes close to it at much lower expense. See Recommended effort levels for Claude Opus 5.5.

At an offered level, Claude Opus 5.5 tends to believe more per turn than Claude Opus 5, specifically at xhigh and maxIf you keep the effort worth you set for Claude Opus 5, anticipate longer turns and more output tokens. 3 changes assistance:

• Set max_tokens high enough to leave space for the design’s believing tokens and the reply. Believing counts towards max_tokens even when believing material isn’t gone back to you, so a limitation sized for Claude Opus 5 with believing off can cut replies off. For the long turns that agentic coding can produce, a max_tokens of 128,000, the design’s optimum, has actually worked well in Anthropic’s screening.
• Reserve xhigh and max for work where you’ve determined a quality gain.
• To get less thinking, lower the effort level. Reducing effort lowers thinking, and with it cost and latency, more dependably than timely guidelines do.

Altering the high-level effort worth in between demands revokes the timely cache. To run specific turns at a various level, utilize a per-message effort modification (beta) rather, which keeps the cache.

Triggers composed for believing handicapped

Claude Opus 5 accepts thinking: at high effort or listed below; Claude Opus 5.5 does not, and the migration guide covers the demand modification. If your Claude Opus 5 combination kept up believing handicapped, 4 modifications choose it:

• Start at low effort and procedure. At low the design keeps its thinking brief. How typically it avoids believing completely depends upon your triggers, so determine latency and quality by yourself traffic and transfer to medium if quality drops. If time to very first token still matters after that, a system timely line such as “Answer directly without deliberating.” can lower believing even more; determine quality when you include it, since less thinking can reduce it.
• Eliminate guidelines that stood in for believing. If your timely asked the design to draw up its thinking in the action as an alternative for believing, get rid of that guideline and check out the thinking from summed up thinking obstructs rather (display screen: "summarized"; a timely that presses the design to replicate its thinking in the reaction text can be decreased with the reasoning_extraction rejection classification.
• Re-test the thinking-disabled mitigations. Keeping up believing handicapped advises a combined guideline (authorization to speak before a tool call, what to do when no tool fits, no internal tags) and getting rid of any guideline that informs the design not to believe. Both address artifacts that appear on Claude Opus 5 just when believing is handicapped. With believing constantly on, examine whether you still require the direction, and eliminate the no-thinking guideline in either case.
• Check out the reaction by block type. Inspect each block’s type rather of presuming the very first material block is text: an action might or might not start with a thinking block, whose thinking field is empty under the default display screen: "omitted"

Ignored agentic runs

On long jobs with a number of parts, Claude Opus 5.5 keeps the user upgraded as it works, and a few of those updates end the turn with text instead of a tool call (stop_reason: "end_turn". An ignored representative loop that deals with such a turn as completion of the job stops running there. A couple of harness and timely modifications assist it keep running.

Deal with a text-only end of turn as a report instead of as evidence the job is done. Keep the job’s parts in a list the design updates, such as a to-do tool or a file. If a turn ends with products still open and no blocker specified, send out a brief user message calling them, like the following one. You can likewise specify the conclusion condition in advance and have a different, smaller sized design examine the discussion versus it at each end of turn, returning its factor as the next user message when the condition isn’t satisfied. In any case, stop after 2 or 3 automated extensions on the exact same job instead of duplicating them forever, so that a run that is truly stuck ends and can be examined.

If something the design began is still running, such as a background command or a subagent, do not deal with the job as done yet: wait on it to complete and return its output to the design as the next user message.

A system timely addition can likewise make these early stops less regular. Claude Opus 5.5 is responsive to guidelines that call the particular sort of early stop you desire it to prevent, such as ending the turn with a summary that reveals the next action rather of taking it. It likewise assists to call the stops you do desire, for instance when no work can advance without the user’s input.

The following paragraph is one example of such an addition, composed for representatives that run totally ignored, where you desire the design to keep working instead of stop to report. Treat it as a beginning point: you may require to adjust it for your own application. Include it at the end of your system trigger from the very first demand of the session: including it partway through modifications the system timely and revokes the discussion’s earlier thinking blocks (see Preserved thinking). Due to the fact that it informs the design to put status notes in the very same message as its next tool call, those notes show up in between tool calls as development updates, whose text returns empty at the default thinking.display; set display screen: "updates" to get a summary of each (see User-facing development updates). With this addition the design continues where it would otherwise have actually stopped to sign in, so keep your own verification action for dangerous or irreparable actions, and leave the addition out of human-in-the-loop applications, where somebody exists to address. Anticipate rather more tool calls and output tokens per job.

Protect rejections

Claude Opus 5.5 runs security classifiers, consisting of for biology, cybersecurity, and thinking extraction.

• Biology: The biology safeguards are the exact same as Claude Fable 5.1’s and are brand-new if you’re originating from Claude Opus 5. Daily health and academic concerns are untouched. If the biology classifier obstructs of your company’s life sciences work, use to the Life Sciences Verification Program
• Cybersecurity: Discovering vulnerabilities in source code is permitted. High-risk dual-use cybersecurity activities are not.
• Thinking extraction: Demands that press the design to replicate its internal thinking in the reaction text can be decreased with the reasoning_extraction classification, which is brand-new if you’re originating from Claude Opus 5. If your triggers ask the design to draw up its thinking in the action, get rid of those guidelines, set display screen: "summarized"and check out the summed up thinking from the thinking obstructs rather; see Prompts composed for believing handicapped.

A classifier decrease gets here as a regular action with stop_reason: "refusal" and a stop_details things calling the classification. You can have the demand retried instantly on a fallback design, other than for reasoning_extraction decreases, which server-side fallback go back to you rather of retrying; see Refusals and alternative.

User-facing development updates

In between tool calls, Claude Opus 5.5 composes brief user-facing development updates: what it simply discovered and what it’s doing next. 4 levers manage what your users see.

Examine that your customer gets them: on Claude Opus 5.5 these notes come back as progress-update thinking blocks instead of text blocks, and their text is empty at the default thinking.displayso a customer that renders just text blocks can look quiet throughout a long agentic turn. Set screen: "updates" (beta, thinking-display-updates-2026-08-18 header) to get a brief summary of each note; the migration guide demonstrates how to render them.

Second, if the design might require to hand the user something verbatim partway through a long turn, such as a code bit, offer it a basic tool for sending out the user a message and inform it to schedule the tool for that material. State the tool in tools from the very first demand of the session: including it to tools later on modifies the discussion’s prefix and revokes earlier thinking blocks (see Preserved thinking).

Third, if you desire more regular or foreseeable updates, such as a one-line declaration of intent before the very first tool call and a brief wrap-up at the end, state so in the system timely; the design is responsive to such guidelines. This assists most in human-in-the-loop work.

4th, if long tool-calling turns still go peaceful for longer than you desire, have your harness request an upgrade. With screen: "updates" set (the very first lever), count successive tool-calling actions that provide the user absolutely nothing to check out: no text block and no progress-update text. After a number of in a row (5, for instance), add a suggestion like the following one after the most recent tool outcomes, as a turn-scoped system message (clear_at: "next_user_message"; beta, mid-conversation-system-clear-at-2026-08-21 header). If the turn remains peaceful, stop after 2 or 3 tips instead of sending out more. Due to the fact that each tip is added and left in location, instead of placed for one demand and erased on the next, the timely cache keeps matching and the thinking obstructs that follow it remain legitimate. In Anthropic’s screening on agentic coding jobs, this approximately cut in half the share of jobs with a long quiet stretch, without any quantifiable modification in expense.

Check out context in multi-app workflows

In workflow automation throughout numerous linked apps, such as e-mail, files, spreadsheets, and CRM records, the details a job depends upon frequently sits someplace the demand does not clearly point out: for instance, a policy in an old e-mail thread, a guideline on another spreadsheet tab, or a note on a consumer record. Claude Opus 5.5 tends to get to work rapidly, and on loosely defined jobs it assists to inform the design to browse the pertinent sources before acting. If your representative works throughout numerous apps on jobs like these, one sentence in the system timely makes it take a look around before it alters anything:

In Anthropic’s screening on multi-app automation jobs, Claude Opus 5.5 finished visibly more of them properly with this guideline, at both medium and max effort, at the expense of somewhat more tool calls and tokens. Since it informs the design to act upon what it discovers, keep untrusted material out of the records it browses.

Time signals for multiagent harnesses

Claude Opus 5.5 pays very close attention to details about elapsed time, and in a multiagent setup, for instance a lead representative that delegates to subagents, you can utilize that to accelerate the resolve much better parallelization. If you can approximate for how long the job needs to take, provide the design a time budget plan: have your harness include a brief line at the end of each message it returns to the design providing the elapsed time versus that spending plan, in seconds, for instance elapsed 340s / 1200sThe design paces its work to complete inside the spending plan and typically ends up well before it, so set the spending plan rather above the time you in fact desire invested and tune it on a sample of your own jobs. If you can’t anticipate a practical spending plan, reveal the elapsed time alone and include one sentence to the system timely:

In Anthropic’s examinations of little representative groups on research study jobs, both signals made groups complete earlier than a single representative working without them. Groups offered a spending plan kept response quality similar to the single representative’s while ending up substantially earlier. A tighter budget plan has a various impact from a lower effort setting: reducing effort minimizes the work itself, whereas a budget plan mainly keeps more representatives operating in parallel. The budget plan is advisory and absolutely nothing stops the design at the limitation, so if you require a difficult stop, keep your own timeout. Inspect respond to quality on your own jobs, due to the fact that under time pressure the design may browse and confirm a little less.

Believing guidelines in chat system triggers

In chat applications, if your system timely includes directions that inform Claude to believe thoroughly before responding to, think about eliminating them for Claude Opus 5.5. The design chooses for itself just how much to believe, and effort is the primary control. In Anthropic’s screening in a chat item, eliminating such a line made replies begin quicker, without any clear decrease in the quality of the reply.

In multi-turn chat, Claude Opus 5.5 in some cases returns over an earlier response while it thinks of a brand-new message, even a brief follow-up, which includes thinking and latency on later turns. If you would rather the design reward earlier responses as settled, include 2 sentences at the end of the system timely:

In Anthropic’s screening this decreased thinking on follow-up turns and made replies begin quicker without impacting quality. Leave it out where you desire the design to keep re-examining its earlier work, for instance in long analyses, or in agentic jobs where a later action can expose an error in an earlier one. The guideline might likewise make the design less most likely to mention an error in an earlier response by itself, so if that matters for your application, test for it before embracing the direction.

Mark pasted text in user messages

Claude Opus 5.5 withstands indirect timely injection, indicating guidelines that show up through tool outcomes, websites, and on-screen or web browser material, much better than any earlier Opus design. With the ideal context it is likewise robust versus directions inside material a user copied into their message from somewhere else, such as an e-mail or a websites. To get that habits, mark which text is the user’s own and which was pasted from elsewhere. Wrap each pasted block in an opening and a closing tag that both bring the exact same brief random ID, created by your application, with each tag on its own line:

Include this note to your system timely:

This can make the design a little more mindful sometimes, so determine the impact by yourself jobs. The tags appear text and can be mimicked, so treat this as one guardrail together with other prompt-injection defenses.

Due To The Fact That Claude Opus 5.5 checks out charts, diagrams, and screenshots substantially more specifically than Claude Opus 5 without tools (see Capabilities pertinent to triggering), re-test whether you still require scaffolding you constructed for visual inputs on earlier designs. For the densest inputs, 2 things still include precision. Higher-resolution images assist, many of all for inputs like technical illustrations. Do image-processing tools: run the design as a representative with access to a container that holds the raw images and has libraries such as PIL and OpenCV set up, so that it can crop, zoom, determine, and confirm its work. If a container is excessive overhead, a cropping tool alone still assists; the crop tool dish has a working meaning. The design utilizes these tools better at greater effort levels. Without tools, raising effort enhances its reading of technical illustrations however does little for charts.

Frontend style defaults

Requested for frontend work without style instructions, Claude Opus 5.5 draws on a couple of default designs, and a basic direction such as “avoid a generic AI look” primarily swaps one default for another. It reacts well to guidelines that call particular patterns to prevent, as in the copying. Work iteratively: examine which designs the very first outcome utilized rather, and extend the list if required.