How to prompt 📝Claude Opus 5.5 comes down to five changes from Claude Opus 5: choose an effort level instead of disabling thinking, treat a text-only turn ending as a report rather than completion, render progress updates from thinking blocks, name the specific defaults to avoid, and mark text the user pasted.

This guide condenses 📝Anthropic's official Prompting Claude Opus 5.5 guide and its What's new notes into symptom-first answers, with our own operating notes. Anthropic states that existing Opus 5 prompts should perform well on Opus 5.5 without changes; the adjustments below address the behaviors that do differ. Every section stands alone, so a reader or agent can jump straight to the symptom they are seeing.

Quick Diagnosis

Turns run longer and cost more than on Opus 5 — recalibrate effort, starting at medium
Requests fail with a 400 error about thinking — remove thinking.disabled and manual budget_tokens
Requests fail with a 400 error about tool_choice — use auto with strict tool use instead of forced tools
An unattended agent stops partway through a task — treat text-only turn endings as reports and continue
Long agentic turns look silent to users — set thinking.display to "updates"
Responses return stop_reason: "refusal" — handle safeguard categories and configure fallback
Chat replies start slowly — remove "think carefully" instructions from the system prompt
The model obeys instructions inside pasted text — wrap pasted content in ID-tagged blocks
Agents miss context across connected apps — tell the model to explore before acting
Frontend output looks generic — name the specific styles to avoid

What effort level should I use for Claude Opus 5.5?

Start at medium, which is the default on Opus 5.5, and set it explicitly. Anthropic reports that Opus 5.5 at medium matches or exceeds Opus 5 at high on coding and knowledge-work evaluations, and that low comes close on several coding evaluations at much lower cost.

Effort names do not map to the same amount of thinking across models, so an effort setting carried over from Opus 5 produces longer turns and more output tokens. Reserve xhigh and max for work where a quality gain has been measured, and set max_tokens high enough to hold thinking plus the reply; Anthropic found 128,000 works well for long agentic coding turns. Changing top-level effort between requests invalidates the prompt cache, so use per-message effort (beta) to vary it within a conversation.

Can I still disable thinking on Claude Opus 5.5?

No. Thinking is always on, and thinking: {"type": "disabled"} or a manual budget_tokens returns a 400 error. Integrations that disabled thinking on Opus 5 should start at low effort and measure quality and latency.

Remove prompt instructions that asked the model to write out its reasoning as a substitute for thinking. On Opus 5.5 those instructions can trigger a reasoning_extraction refusal; read reasoning from summarized thinking blocks (display: "summarized") instead. Parse responses by content block type rather than position, because a response may begin with an empty thinking block.

Why does my Opus 5.5 agent stop partway through a long task?

Opus 5.5 reports progress as it works, and some of those reports end the turn with text instead of a tool call. An unattended loop that treats stop_reason: "end_turn" as task completion stops there.

Keep the task's parts in a checklist the model updates, and when a turn ends with open items and no stated blocker, send a short continuation message. Anthropic's 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." Cap automatic continuations at two or three per task so a genuinely stuck run ends for review.

For fully unattended agents, Anthropic also publishes a standing system-prompt instruction that names four unwanted early stops — a summary that announces the next step without taking it, an offer to continue, a list of non-blocking decisions, and a self-declared milestone report. Add it from the first request only, keep your own confirmation step for risky actions, and leave it out of human-in-the-loop applications.

Why do long agentic turns look silent?

On Opus 5.5, the notes the model writes between tool calls arrive as thinking blocks, not text blocks, and their text is empty at the default display: "omitted". A client that renders only text blocks goes quiet with no error.

Set thinking.display to "updates" (beta header thinking-display-updates-2026-08-18) to receive them. Ask in the system prompt for a one-line statement of intent before the first tool call and a short recap at the end. If a turn still runs five or more tool calls without user-visible text, append a turn-scoped reminder such as "The user hasn't heard from you in a while — say in a few words what you're doing, then continue." Anthropic reports this roughly halved long silent stretches with no measurable cost change.

Why does Claude Opus 5.5 refuse some requests?

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

The biology classifier is new relative to Opus 5; organizations doing life-sciences work can apply to the 📝Anthropic Life Sciences Verification Program. Finding vulnerabilities in source code is allowed, while high-risk dual-use cybersecurity activity is not. Configure server-side fallback with fallbacks: "default" to retry declined requests on the model Anthropic recommends, except reasoning_extraction declines, which fallback returns rather than retries.

Should chat system prompts still tell 📝Claude to think carefully?

No. Anthropic found that removing "think carefully before answering" lines from a chat system prompt made Opus 5.5 replies start sooner with no clear quality loss, because the model decides its own thinking depth and effort is the control.

In multi-turn chat, Opus 5.5 sometimes re-examines earlier answers on short follow-ups. Where earlier answers should be treated as settled, Anthropic suggests ending the system prompt with: "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." Leave it out of long analyses and agentic work, where later steps should be able to catch earlier mistakes.

How do I stop Opus 5.5 from following instructions in pasted text?

Wrap each block the user pasted in opening and closing <pasted_content> tags that carry the same short random ID generated by the application, and add a system-prompt note explaining that the text came from elsewhere and should be followed only where the user's own message asks.

Opus 5.5 resists indirect prompt injection better than any earlier Opus model, and this marking extends that resistance to copied emails and web pages. The tags are plain text and can be imitated, so treat them as one guardrail among several.

How do I get better results across connected apps?

Tell the model to look before it acts. Opus 5.5 gets to work quickly, and on loosely specified multi-app tasks it can miss a policy in an old email or a rule on another spreadsheet tab. Anthropic's one-line addition — "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." — completed noticeably more tasks correctly at both medium and max effort. Keep untrusted content out of the sources it searches.

How do I make a multiagent team finish sooner?

Give the lead agent a time signal. Opus 5.5 paces its work against elapsed time, so a harness line such as elapsed 340s / 1200s appended to each message lets it parallelize subagents to finish inside the budget. Without a predictable budget, show elapsed time alone and state that earlier correct results are better. The budget is advisory; keep a hard timeout of your own.

How do I get less generic frontend designs from Opus 5.5?

Name the specific patterns to avoid rather than asking it to "avoid a generic AI look," which only swaps one default for another. Anthropic's example excludes cream backgrounds, italic accent words in headlines, numbered section labels, monospace labels, and pill-shaped buttons. Check which styles the first result used and extend the list.

Does Opus 5.5 still need vision tooling?

Less than earlier models. Opus 5.5 reads dense charts more accurately at its lowest effort than Opus 5 did at its highest, so re-test scaffolding built for earlier models. For the densest inputs, such as technical drawings, higher-resolution images and a crop or zoom tool still add accuracy.

Operating Rules

Standing rules for agents running on Claude Opus 5.5, as of September 28, 2026:

Effort — set medium explicitly; drop to low for latency-bound work; use xhigh or max only with a measured gain
max_tokens — size for thinking plus reply; use 128,000 for long agentic turns
Thinking — never send disabled or budget_tokens; never ask the model to write out its reasoning in the reply
Tools — use tool_choice: auto with strict tool use; declare every tool from the first request
Conversation history — keep it append-only; change instructions with mid-conversation system messages
Turn endings — treat text-only endings as reports; continue on open checklist items, at most three times
Refusals — branch on stop_reason: "refusal"; configure fallback for cyber and bio categories
Pasted text — wrap user-pasted content in ID-matched <pasted_content> tags

📝Claude Opus 5.5 — specs, pricing, and capabilities
📝Claude Opus 5 vs Claude Opus 5.5 — what changed from the predecessor
📝Claude Model Glossary — every Claude model by tier
📝Prompt Injection — the attack the pasted-content convention guards against
📝Context Engineering Rules for Claude 5 Generation — context practices across the Claude 5 models
📝Claude Code — the agentic coding surface where these patterns apply
📝Claude Platform Glossary — API terms used throughout this guide

We run BotBrian on Opus 5.5 inside Claude Code, and two of these patterns already show up in our setup: the Claude Code harness wraps pasted text in <pasted_content> tags, and our global instructions ask the agent to state its next move in one line before tool work — the same lever as Anthropic's progress-update guidance. We will add field observations here as they accumulate.