Anthropic outlines how to get better results from Claude Opus 5.5

Anthropic is advising Claude users to give its Opus 5.5 model clearer goals, then allow it to work with less frequent intervention. The company’s guidance focuses on long tasks, reviewing results, handling documents and images, and avoiding prompt habits that may no longer be useful. Addy Osmani writes for Anthropic that Opus 5.5 is designed to continue through longer, multi-step assignments and to state plainly what it has completed or still needs from a user.

The central recommendation is simple: describe the full assignment and define what completion means. Instead of breaking a project into many small requests, users should set a measurable finish line. For example, a content team could ask Claude to review a presentation and consider the task complete only after it has checked dates, figures, names, and internal contradictions.

Anthropic also advises users to specify when the model should pause for input. This matters most in Claude Code, the company’s tool for working inside software projects, but the principle also applies to other workflows. Users can ask the model to continue on routine work while stopping for decisions, missing information, or actions that are difficult to reverse.

Clear requests instead of “think carefully” prompts

One of the more notable recommendations is to remove instructions such as “think carefully” or “think step by step.” According to Anthropic, Opus 5.5 already reasons before responding and determines the needed level of effort itself. The company says removing such language can make responses start faster without a clear loss in quality.

For straightforward questions, users can instead ask for a direct answer. The guidance distinguishes this from asking the model to reveal its private internal reasoning. Anthropic says requests to reproduce that reasoning may be declined. A better alternative is to request a brief explanation of the choice or a concise summary of the supporting evidence.

Visual and creative tasks need similarly concrete direction. When requesting a webpage, slide, or other design, Anthropic recommends naming styles to avoid rather than relying on broad instructions such as “make it non-generic.” A useful request might exclude specific elements, such as off-white backgrounds, italicized headline accents, numbered section labels, or pill-shaped buttons. Users can then refine the instruction after seeing the first result.

Use the original materials and check the handoff

For work involving charts, screenshots, diagrams, or slides, Anthropic recommends attaching the original image rather than retyping its contents. The company says Opus 5.5 has improved at interpreting visual relationships, including arrows in diagrams, changes between versions, and time information in calendars.

The model can also check lengthy documents for discrepancies. Users should ask it to quote each issue and identify where it appears. For research tasks, Anthropic recommends explicitly requesting a section on what could not be verified and where the model looked. That can make uncertainty easier to spot before material is published or shared.

When a long task ends, the company advises reading the model’s outstanding requests first. Those may include approvals, choices, or missing details that prevent a final step. Users can also define a standard closing format, such as “Blocked on me,” “Changed,” and “Found.”

Finally, Anthropic suggests asking Opus 5.5 to review a document, a set of changes, or a draft before human review. The most useful review prompts set a high bar: ask for only issues serious enough to block approval, along with the location, the problem, and a way to demonstrate it.