AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Anthropic has published guidance for using Opus 5.5 in Claude apps and Claude Code, focused on handing over well-defined tasks, steering longer runs and checking results. The recommendations are based on the company’s own product guidance and testing; the source does not provide independent benchmarks or detailed performance data.

Anthropic has published a guide to getting more out of Opus 5.5 in Claude and Claude Code, advising users to define a task’s finish line, give the model room to work, and set clear rules for when it should stop and ask for input. The recommendations matter because the company says the model handles longer, multi-step tasks and behaves differently from earlier Opus models.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

The guide’s central recommendation is to give Claude the whole task in one message, including what completion means and the conditions that should trigger a pause. For a code migration, for example, Anthropic suggests specifying that all endpoints must use a replacement client, the old client must be removed, and the test suite must pass. It advises users to let the model continue until that stated finish line, rather than breaking the work into repeated prompts.

Anthropic also says users can remove instructions such as “think carefully” or “think step by step.” The company says Opus 5.5 already thinks before replying and decides how much effort to use. In its testing in a chat product, removing a “think carefully” instruction made replies start sooner, with no clear drop in quality. The guide suggests asking for a direct answer when that is what the user wants, and adjusting the effort setting in Claude Code when appropriate.

For longer Claude Code runs, Anthropic recommends spelling out when the model should continue and when it needs human approval. A project’s CLAUDE.md file can instruct it to keep going when a step does not require input, while stopping before destructive actions such as deleting data or force-pushing. The company also suggests keeping a task checklist in a file such as TASKS.md, asking the model to divide large audits across subagents and verify their evidence, and sending follow-up instructions during a run rather than restarting it.

At a glance
announcementWhen: Published in Anthropic’s Opus 5.5 guida…
The developmentAnthropic published a practical guide to prompting and supervising Opus 5.5 in Claude and Claude Code.

Making Longer Claude Code Runs Easier to Manage

The advice addresses a practical trade-off in using a model for extended work: fewer interruptions can save time, but they also give the model more opportunity to act before a person reviews its progress. Explicit completion criteria and stop conditions help users delegate more of a multi-step job without treating every action as automatically safe.

Anthropic’s recommendations also put emphasis on reviewing evidence and preserving oversight. A checklist can make progress easier to inspect when a long conversation is summarized, while asking subagents to report evidence gives the user something to check rather than relying only on a final conclusion. These are workflow suggestions, not a guarantee that tasks will complete correctly or that independent checks are unnecessary.

The guide is relevant to developers and other Claude users who handle repository-wide changes, audits or design work. It also describes a change in prompting habits: Anthropic says users may get better results from clear objectives and explicit constraints than from adding generic instructions to think harder or avoid a generic design.

Products Worth Considering

Apple AirPods Pro (2nd Generation) Wireless Earbuds with MagSafe Charging Case (Lightning)

Apple AirPods Pro (2nd Generation) Wireless Earbuds with MagSafe Charging Case (Lightning)

As an affiliate, we earn on qualifying purchases.

What Anthropic Says Changed in Opus 5.5

Anthropic says Opus 5.5 works for longer without supervision than Opus 5 and performs better on multi-step work, including carrying a change through a large repository until tests pass. The company reports that early testers ran lengthy coding tasks for hours with little oversight, and that testers used coordinated subagents for audits and migrations. These are claims in Anthropic’s own guidance; the supplied report does not include a benchmark, sample size or independent evaluation.

The guidance covers both model behavior and user controls. Anthropic says Opus 5.5 tells users what it has done and thinks before each reply, but it may also pause to summarize, offer to continue or present choices that do not actually block progress. Its suggested CLAUDE.md instructions aim to reduce those unnecessary stops while preserving pauses for actions that require approval.

For visual and design tasks, the company says a broad request to avoid a generic look may simply lead the model to another familiar default. It recommends naming specific patterns to avoid, such as certain background colors, heading treatments or button shapes, and refining the list after seeing the output. This keeps the advice focused on concrete constraints rather than relying on broad stylistic language.

“Stop and ask only when you can’t continue without me, or before anything destructive.”

— Anthropic’s suggested CLAUDE.md instruction

What the Guide Does Not Establish

The report is a vendor-authored usage guide, not an independent assessment of Opus 5.5. It does not give benchmark results, a quantified comparison with Opus 5, details of the early testers’ tasks, or a measured success rate for long autonomous runs. Anthropic’s statements about improved multi-step performance and hours of work with little oversight should be understood as company claims.

The guide also does not establish that fewer interruptions or subagent use will improve results in every project. A model may still make mistakes, miss requirements or misread a repository. The source advises checking results, but the supplied material ends during its section on reviewing completed work, so it does not provide the full set of result-checking recommendations. The publication date and any rollout or availability details are not stated in the provided report.

How Users Can Apply the Advice

Users can test the recommendations on a bounded task: state the intended result, list what counts as completion, and specify when Claude should pause for input. In Claude Code, they can put project-specific stop rules in CLAUDE.md, keep a durable checklist for extended work, and retain permission prompts for destructive commands.

For work divided among subagents, Anthropic recommends asking the model to check each report’s evidence before combining the findings. Users can then inspect the output and test results themselves. Anthropic’s source does not announce a separate product change or give a future milestone; it presents these steps as ways to adapt existing workflows to Opus 5.5.

Key Questions

What is Anthropic’s main recommendation for using Opus 5.5?

Give it the full task in one message, define what completion looks like and say when it should stop for input. Anthropic recommends letting the model work toward that finish line when the task does not require an earlier decision.

Should users tell Opus 5.5 to think step by step?

Anthropic advises removing generic prompts such as “think carefully” because it says Opus 5.5 already thinks before replying. The company reports that removing one such instruction made responses start sooner in its chat-product testing, with no clear drop in quality.

How can users steer long Claude Code tasks?

Anthropic suggests adding continue-or-stop rules to CLAUDE.md, using a file-based task checklist and sending follow-up instructions while a run is in progress. It says users should preserve approval requirements for destructive or hard-to-reverse actions.

Does Anthropic’s guide prove Opus 5.5 is more reliable than earlier models?

No. The guide says Opus 5.5 performs better at multi-step work than prior Opus models, but the supplied report provides no benchmark figures or independent evaluation. Its performance statements are Anthropic’s claims.

Source: hn

You May Also Like

What The Science Paper Says About Microsoft’s AI And Doctors

A peer-reviewed Science paper reports high diagnostic scores for Microsoft’s MAI-DxO, but its results come from simulated cases, not patient care.

Unveiling Zhang Yiming’s Latest AI Company With A 10 Billion RMB Valuation

ByteDance founder Zhang Yiming has incubated a new AI company valued at 10 billion RMB, signaling ongoing personal investments in China’s AI sector.

Inkling: Our Open-Weights Model

Inkling announces its new open-weights model, enabling broader access and customization for AI applications, marking a significant step in open AI research.

Why Claude Fable 5.1 Is At The Top Of The AI Index — And What To Know About The Cost Line

Claude Fable 5.1 achieves the highest score ever on the AI Intelligence Index, but at a higher cost and with increased verbosity. Here’s what to know.