Anthropic's official prompting guidance for Claude Opus 5.5, published with the model's September 22, 2026 launch — how to calibrate effort now that the default dropped to medium, why thinking is always on, and which chat-prompt lines to delete. Guidance for a text model, not a content generator.
Last verified · 2026-09-28 · by Moe Ameen
The Claude Opus 5.5 prompting guide is Anthropic's official documentation on how to prompt its September 2026 flagship model, published as part of the launch and covered widely the week after. Its framing is unusual: instead of promising better output from the same prompts, it warns that settings carried over from Claude Opus 5 can now run longer and cost more, and it walks through the behavioral changes that make old habits misfire. Anthropic is explicit that existing Opus 5 prompts should still work — the guide is about tuning, not rescue.
The headline changes are about thinking and effort. Opus 5.5's default effort level is medium (Opus 5 defaulted to high), and thinking is now always on — the model rejects requests that try to disable it. Because it tends to think more per turn at any given level, Anthropic advises setting effort explicitly, starting at medium, and testing several levels against your own evaluations rather than assuming the level names carry the same meaning across models; it also says to budget max_tokens generously (up to the model's 128,000 maximum on long agentic turns) because thinking counts toward that limit. For chat apps, it recommends deleting "think carefully" style instructions, since the model self-regulates thinking and removing the line made replies start sooner with no clear quality drop.
The rest is production-grade advice: keeping unattended agents running past a text-only end of turn, reading progress updates that now arrive as thinking blocks, giving multi-agent harnesses an elapsed-time budget, wrapping user-pasted text in tagged blocks to resist prompt injection, using tools for dense charts and diagrams, and naming specific frontend styles to avoid rather than asking it to "avoid a generic AI look." It's engineering guidance written for people building on the API — verify any figure against Anthropic's docs, since the advice is versioned to the model and can change.
Most of this guide is plumbing a creator will never touch — but two of its techniques transfer directly to the one prompting a creator does do: writing the brief that drives their content. First, the guide's best frontend tip is that "avoid a generic AI look" does nothing, while naming the specific patterns to avoid works. That's exactly how you write a good [Persona Brief](/glossary/persona-brief) in [Kompozy](/): don't say "sound human," list the specific words, openings, and clichés to ban. Second, the guide tells agents to explore a source broadly before acting instead of jumping on the obvious. Applied to content, that means feeding Kompozy a rich, complete source — the full transcript, the whole report — rather than a thin summary, so the copy it generates has real material to draw from. Take those two habits, write one genuinely sharp brief and source, and you've done the only prompting that matters for content.
Then Kompozy does what no prompt can: it turns that brief into finished, published posts. Its Text Posts, Blog Articles, and Newsletters run on this class of Claude and OpenAI model — with the effort and thinking calibration the guide describes handled internally, so you never open a docs page to tune it — and it generates the formats a text model can't: talking-head [Persona Shorts](/glossary/persona-shorts) and HeyGen avatar video, brand-exact Carousels and Quote Graphics rendered through [HyperFrames](/glossary/hyperframes), Photo Posts, and Infographics, all held to one voice. [Autopilot](/glossary/autopilot) and a per-post review pipeline then caption, reframe to 9:16, 1:1, and 16:9, schedule, and publish across the eight social platforms plus a blog and a Mailchimp newsletter. The guide teaches you to get one great answer out of the model; Kompozy turns one great brief into a week of content, everywhere.
For chat users, the main practical change is to delete "think carefully" style instructions from saved prompts and custom instructions. Opus 5.5 sets its own thinking on every reply, so those lines can slow the first reply without improving quality. The deeper advice — effort calibration, token budgets, agent stop conditions — is aimed at developers building on the API, not chat users.
Medium. That's one step below Claude Opus 5's high default. Anthropic recommends setting effort explicitly and re-testing several levels on your own evaluations, because the level names don't map to the same amount of thinking across models — at medium, Opus 5.5 matched or beat Opus 5 at high on coding and knowledge-work tasks in its testing.
Two of its techniques transfer to content: name the specific patterns to avoid rather than saying "sound human," and give the model a rich source instead of a thin one. Both apply directly to writing a Kompozy Persona Brief and choosing a source. The rest is engineering plumbing, and none of it makes the model output video, images, or scheduled posts.
Not for content. Kompozy generates its copy with this class of Claude and OpenAI model and manages the effort and thinking settings the guide describes internally, so you don't calibrate anything. You'd only need the guide if you were separately building your own app or agent on the raw Claude API.