Claude Opus 5.5 Prompting: 6 Habits Anthropic Says to Drop

Anthropic says the phrase “think hard” now makes Claude Opus 5.5 slower without making it better. In its testing, removing that line made replies start sooner with no clear drop in quality. For teams and heavy users who built prompt libraries on older Claude models, that means some of their habits now cost them speed.

The Framework

The Decoder reports that Opus 5.5, now the default model on Pro, Max, Ultra and Team plans, works differently from its predecessors. Anthropic’s own guide, Using Claude Opus 5.5 as your daily driver for knowledge work, says the principle behind it plainly: “Each new model is a reason to retest what you hand to Claude and to drop prompting habits built for older models.” Most of the changes come down to two ideas. The model now decides how much to reason, and it can carry a whole task from one well-specified request.

1. Delete “Think Hard” and “Think Carefully”

Thinking is always on for Opus 5.5, and the model sets its own reasoning depth from the request and the effort level. The Decoder notes these phrases date from early reasoning models, before adaptive reasoning was standard.

2. Start at Medium Effort

Medium is the default. Anthropic says that on its coding and knowledge-work evaluations, Opus 5.5 at Medium matched or beat Opus 5 at High, and that higher settings think longer and cost more without always producing better work. An early tester at Factory, Zimu Li, said it “matched Opus 5 on high effort, while using 20 to 25% fewer output tokens.” Move to High when Medium stalls.

3. Say What Done Looks Like in the First Message

Rather than feeding instructions one at a time, Anthropic recommends describing the goal, the finished output and when the model should stop and ask. It says Opus 5.5 carries long, multipart work through to the end better than Opus 5 did.

4. Attach the Chart Instead of Retyping It

Workarounds such as transcribing a chart’s values may no longer be needed. Anthropic says that even at its lowest effort setting, Opus 5.5 read values off dense charts more accurately than Opus 5 did at its highest.

5. Ask for the Finished File

If the output is a spreadsheet or document, ask for that rather than an outline. Anthropic says Opus 5.5 can often go from request to finished file in one pass.

6. Name the Styles You Don’t Want

For design work, a vague ban on “a generic look” tends to swap one default style for another. Anthropic suggests naming specific patterns, such as a cream background or pill-shaped buttons. It also recommends connecting Claude to work files through connectors and supplying a house style guide.

Where It Doesn’t Apply

The model-choice advice has a price attached. Anthropic lists Sonnet 5.5 at $2 / $10 per million input and output tokens and Opus 5.5 at $4 / $20. Its case for defaulting to Opus is that a more capable model can finish a task doing less work, which, according to its cost guidance, can offset much of the higher per-token price. It adds that pushing Sonnet to Extra high or Max effort brings cost per task close to Opus. On subscription plans, higher-priced models use limits faster, and on Team and Enterprise plans an owner has to enable a connector before members can use it.

The benchmark claims are Anthropic’s own internal tests, so the gains are worth checking on a team’s own work. Stack Archive’s coverage of the Opus 5.5 and GPT-6 Sol and Luna launches covers the pricing context these defaults sit in.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.