Anthropic tells Claude users to drop 'think hard' prompts: Opus 5.5 works better without old habits

Anthropic says users are slowing down their own model. In guidance published alongside the rollout of Claude Opus 5.5, now the default on Pro, Max, Ultra and Team plans, the company told customers to drop prompting habits carried over from older model generations — starting with the phrase "think hard."

The advice matters because prompting conventions spread quickly through the AI user community, and they calcify. Instructions that once squeezed extra performance out of reasoning models roughly a year and a half ago now add latency and complexity without adding anything back, according to Anthropic. The company's position is direct: Opus 5.5 works differently from its predecessors, and users who have not updated their habits are holding it back.

Why 'think hard' no longer helps

Topping Anthropic's list is a call to cut phrases like "think hard" or "think carefully" from prompts. The reason is architectural rather than stylistic. The model always reasons on its own and scales its reasoning depth to the task at hand and to the effort level the user has selected, the company says. Adaptive reasoning — the model's ability to decide for itself how much deliberation a task warrants — is now standard behavior rather than something users have to coax out.

In Anthropic's tests, responses without these phrases started faster, with no measurable drop in quality. The phrases date back to the early days of reasoning models, before adaptive reasoning existed in its current form. "Back then," meaning about a year and a half ago, users could sometimes coax more performance out of a model or steer the router — the component that decides how much computation to spend on a request — one way or another. That era is over, per Anthropic, and the phrases are now dead weight.

The instruction carries a broader implication. Prompting advice that circulates in tutorials, internal company playbooks and social media threads has a shelf life, and a substantial amount of it is tied to model behavior that no longer exists. Anthropic is effectively telling its customers to audit their templates.

Start at Medium effort, not High

Anthropic's second recommendation concerns the effort setting, a parameter users can change in the model menu next to the send button. The company recommends starting at "Medium," which is the default for Opus 5.5.

The supporting number is the strongest in the guidance: in internal tests, Opus 5.5 running at Medium matched the performance of its predecessor, Opus 5, running at High. Higher settings cost more and do not always produce better results, according to the company. Anthropic's advice is to switch to "High" only when Medium proves insufficient — an inversion of the instinct, common among power users, to reach for the maximum setting first.

For teams that structured their workflows around the older two-model lineup, the economics have also shifted. Anthropic says teams that previously used the cheaper Sonnet for routine work and reserved Opus for complex projects should rethink that split. The company's argument is that a more capable model can finish a task in fewer steps, which can partly offset its higher per-token price. And anyone who regularly pushes Sonnet up to "Extra High" or "Max" effort, Anthropic says, ends up paying about as much as they would with Opus 5.5.

From chatting to setting goals

The third block of guidance concerns conversation structure. Anthropic advises describing the desired result in a single message rather than giving instructions one at a time. That message, the company says, should cover three things:

• The goal of the task.
• The desired output format.
• The conditions under which the model should check back with questions.

Opus 5.5 is better than its predecessor at seeing multi-part tasks through, according to Anthropic. The model can increasingly take over the back-and-forth that users previously handled manually inside the chat — sometimes doing it better — and can effectively prompt itself from a single initial input. In practice, that shifts the user's job from turn-by-turn instruction to upfront specification: state what the finished result looks like, and let the model manage the intermediate steps.

Attach files instead of typing them out

Anthropic also recommends attaching diagrams and tables as images or files rather than transcribing their contents into the prompt. Workarounds that older models required, such as manually typing out chart values, are no longer needed, the company says.

The capability gap here is measurable. Image reading improved significantly in Opus 5.5, according to Anthropic: even at the lowest effort level, the model read values from dense charts more accurately than Opus 5 did at its highest. That is the same pattern as the effort-setting comparison — the new model at a lower tier beating the old model at its ceiling.

The recommendation extends to deliverables. Users who need a finished document or spreadsheet should ask for it directly rather than requesting a draft first and iterating. Opus 5.5 often goes from request to finished product in a single pass, the company says, which makes the draft-then-refine ritual an unnecessary cost in time and tokens.

Name what you do not want

For design tasks, Anthropic's guidance targets a specific and familiar failure mode. The company recommends naming unwanted styles concretely — a cream-colored background, pill-shaped buttons — rather than issuing vague prohibitions against "generic design." According to Anthropic, general instructions only swap one default style for another. A ban on "generic" does not tell the model what to produce instead, so it falls back on whatever default aesthetic it holds; a named, specific exclusion gives it something to actually avoid.

The principle generalizes across task types. Specific, actionable constraints outperform abstract quality demands, because the model can only act on what the instruction actually specifies.

Connectors change the equation

The final recommendation involves connectors, the mechanism by which Claude connects to a user's own files and systems. Users who set up connectors get significantly more tailored results, Anthropic says, because the model can pull in the information it needs on its own rather than depending on the user to paste in every relevant document.

That closes the loop on the through-line of the guidance. Across every recommendation — dropping effort phrases, starting at Medium, specifying the whole task upfront, attaching files, naming concrete constraints — Anthropic is describing the same shift: Opus 5.5 requires less scaffolding from the user, and the scaffolding users still provide can actively hurt. The company's implicit bet is that prompting skill is becoming less about technique and more about clear specification, and customers who keep prompting like it is early 2024 will pay for it in speed, cost and results.