Anthropic tells Opus 5.5 users to delete "think hard" lines

The most useful change to your Opus 5.5 prompts may be a deletion. In Anthropic's chat testing, removing a "think carefully" line made replies start sooner, with no clear drop in quality. That tip opens a new guide on the Claude blog about how to prompt, steer and check Opus 5.5 in Claude apps and Claude Code.

At a glance

• Anthropic's guide says Opus 5.5 behaves differently from Opus 5 in three ways: it works longer on its own, reports plainly what it did, and thinks before every reply.
• For long runs in Claude Code, the guide recommends one message with a finish line, a CLAUDE.md rule on when to stop, and a task list kept in a file that survives context summarization.
• Opus 5.5 is the first Opus model with Fable-level bio and cyber safeguards. Most flagged messages move to an older model, and Anthropic admits the safeguards sometimes flag legitimate work.

If you haven't been following, Opus 5.5 comes after Opus 5, and the guide measures most of its claims against that model. Anthropic says the biggest gains over earlier Opus models are on multi-step work, such as carrying a change through a large repository until the tests pass. Early testers ran long coding tasks for hours with little oversight.

Opus 5.5 always thinks before it replies and decides how much on its own

The guide asks you to remove "think carefully", "think step by step" and similar lines from prompts and saved instructions. Opus 5.5 already thinks before every reply and sets the amount itself, so you don't need to ask. If you want a quick answer to a simple question, write "Answer directly." To change how much it thinks in Claude Code, change the effort setting.

Design work gets a similar fix. When you give no direction, Opus 5.5 falls back on a few default styles, and a vague "avoid a generic look" mostly swaps one default for another. Anthropic says a list of specific patterns works much better. The guide's example bans cream or off-white backgrounds, italic accent words in headings, numbered "01 / 02 / 03" section labels, monospace labels and pill-shaped buttons.

A long run in Claude Code needs a finish line and a rule for when to stop

Give the whole task in one message and say what "done" means. In the guide's migration example, every payment endpoint uses the new client, the old client is deleted and the test suite passes. Claude stops to ask only if a test fails for a reason it can't explain. If you remember something mid-run, type it and press Enter while Claude works. Runs are longer now, so a restart costs more.

Opus 5.5 keeps you posted as it works. On long tasks it sometimes stops to report: it gives a summary that names the next step without taking it, offers to continue, or lists choices that don't block the work. The suggested CLAUDE.md rule says to keep going when a step needs no input. Claude should stop only when it can't continue, or before deleting data, force-pushing or changing anything outside the repository.

Fewer stops mean fewer checkpoints, so the guide says to keep permission prompts on for destructive commands. For large audits and migrations, it suggests giving the work to subagents and checking each one's evidence before accepting it.

A task list in a file survives when Claude Code summarizes older turns

A long run fills the context window, which is the amount of text the model can hold at once. Once it is full, Claude Code summarizes older turns, and early details get compressed. A checklist in a file sits outside that process. It stays whole and shows at a glance what is done and what is left.

Think of a whiteboard next to a meeting whose minutes keep getting shortened: the minutes lose detail, but the whiteboard does not. The guide's prompt is "Keep a checklist in TASKS.md. Tick each item when it's done, and add anything new you find." Then you check progress in the file, not in the scrollback.

At the end of a run, the guide says to read what Claude needs from you first

Anthropic says Opus 5.5 reports more clearly than Opus 5. Its updates and final summary say what it did, what it found and what it needs from you. Read the open decisions and pending approvals first, then the rest. You can set the format in CLAUDE.md, for example with three headings: Blocked on me, Changed, Found.

The guide suggests having Opus 5.5 review a diff before a person does. One early tester said Opus 5.5 at its lowest effort caught more bugs than Opus 5 at high effort, with fewer false alarms. For research, add "Mark anything you couldn't confirm, and say where you looked." This works in a Claude research report and in Claude Code.

In Claude apps, attach charts and screenshots instead of retyping the numbers. Opus 5.5 reads them more accurately than Opus 5, including which boxes an arrow connects. In Anthropic's testing, it caught a date on the wrong weekday in a long planning thread. It also caught a chart that didn't match a deck's numbers.

Flagged messages move to an older model, and you can switch back

In Claude apps and Claude Code, most messages flagged by the new safeguards move to an older model, and your work continues there. Finding security vulnerabilities in source code is allowed. Everyday health and educational questions should still work. The check covers the whole conversation, including files and search results, so a flag can come from earlier content.

In the apps, a notice starting with "Switched to" appears. To return, pick Opus 5.5 in the model picker, or start a new chat, because the earlier message may be flagged again. In Claude Code, run /model to switch back, press Esc twice to edit your last message, and run /feedback if the flag was wrong.

To be asked before a switch, turn off "Switch models when a message is flagged" under Settings, then Capabilities, or change it with /config. A request to reproduce the model's internal reasoning in the reply is one of the flag categories. The guide suggests asking for a three-sentence explanation of the approach instead.

The guide's strongest code-review claim comes from one early tester, and Anthropic admits the safeguards sometimes flag legitimate work. It also says that in long chats Opus 5.5 sometimes goes back over earlier answers, which slows replies. Oddly, asking the model to show its reasoning counts as a flag category, so an old saved instruction could likely move a chat to an older model without warning.

Fast mode and flag tuning

At launch, fast mode for Opus 5.5 is a research preview that you turn on with /fast in Claude Code. It is the same model and the text arrives sooner. It requires extra usage and costs more per token. The guide gives no per-token price and no date for leaving preview. Anthropic says it is tuning the safeguards to cut incorrect flags but gives no timeline.