Delete "think carefully" from your Opus 5.5 prompts

Anthropic's first tip for its newest model is to delete something. In testing in a chat product, removing a "think carefully" line made Opus 5.5 start replying sooner, with no clear drop in quality, because the model already thinks before every reply and decides for itself how much.

That advice opens Claude's guide to getting the most out of Opus 5.5, which also tells you to remove such lines from saved instructions. For a quick answer, say "Answer directly." In Claude Code, you change how much the model thinks with the effort setting.

At a glance

• Anthropic has published a guide for Opus 5.5 in Claude apps and Claude Code. It covers how to prompt the model, steer long runs, check results and recover when a message gets flagged.
• The core pattern is one message with the whole task and a finish line, plus a CLAUDE.md rule saying when to keep going and when to stop before anything destructive.
• New Fable-level bio and cyber safeguards can flag legitimate work, and a flagged message moves the chat or session to an older model until you switch back yourself.

If you missed the release, Opus 5.5 follows Opus 5, and most of the guide lists how the two differ. 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. The ClaudeDevs account boiled the advice down to three first-session tips: hand over the whole task, drop "think carefully", and check what the model needs after a long run.

Early testers let Opus 5.5 run long coding tasks for hours with little oversight

Give the whole task in one message, name the finish line, and say when the model should stop and ask. Anthropic's example is a payment migration. Done means every endpoint uses the new client, the old client is deleted and the test suite passes, and the model stops only if a test fails for a reason it can't explain.

Longer runs make restarts more expensive, so Claude Code lets you type a follow-up and press Enter while Claude works. For audits and migrations across a large codebase, the guide suggests giving each piece to its own subagent and checking each one's evidence before accepting it. Early testers had Opus 5.5 coordinate parallel subagents this way.

For back-and-forth work, where you read each reply before you send the next, Claude Code has fast mode, available for Opus 5.5 at launch as a research preview. You get the same model and the text arrives sooner. Fast mode needs extra usage turned on and costs more per token than standard mode. Type /fast to use it.

A CLAUDE.md rule decides when a long run stops to ask

On a long task, Opus 5.5 sometimes stops just to report. That can be a summary that names the next step without taking it, an offer to continue, or a list of choices that don't block anything. The guide's fix is a CLAUDE.md rule: keep going when a step doesn't need you, put status notes in the same message as the next action, and stop only when blocked.

The rule also says to stop before anything destructive: deleting data, force-pushing, or changing anything outside the repository. Fewer stops mean fewer checkpoints, so Anthropic also says to keep permission prompts on for destructive commands. If a run still ends on "Want me to continue?", reply "continue". For pair programming, write the opposite rule instead: a one-line plan before the work and a recap after.

A TASKS.md checklist 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. When the window fills up, Claude Code summarizes older turns, so early details get compressed. A checklist in a file sits outside that process. It survives intact and shows you at a glance what is done and what is left.

Think of a whiteboard in a shared office compared with someone's memory of last week's meeting. The board doesn't paraphrase. The suggested prompt is to keep a checklist in TASKS.md, tick each item when it's done and add anything new. Then check progress in the file rather than the scrollback.

One early tester says Opus 5.5 at its lowest effort caught more bugs than Opus 5 at high effort

According to the guide, that tester also reported fewer false alarms. Anthropic suggests a review pass before a person reviews the code. The prompt asks for only the problems you would block a merge for, each with the file, the line, why it's wrong and how to show it fails. When a long run ends, first read what Claude is waiting on you for, then the rest of the summary.

In Claude apps, Anthropic says Opus 5.5 reads charts, diagrams and screenshots more accurately than Opus 5, down to 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 the numbers in a deck. For research, ask it to mark anything it couldn't confirm and say where it looked.

Without design direction, Opus 5.5 falls back on a few default styles

A general instruction like "avoid a generic look" mostly swaps one default for another. The guide recommends a specific list. Its example bans cream or off-white backgrounds, italic accent words in headings, numbered "01 / 02 / 03" section labels, monospace labels and pill-shaped buttons. If you dislike what it picks instead, add that to the list and ask again.

Anthropic also says the model's spreadsheets and documents need less editing than Opus 5's, so ask for the finished file rather than an outline. In a long project chat, Opus 5.5 sometimes goes back over earlier answers while it handles a short follow-up. An instruction that earlier answers are settled speeds up replies, but the guide says to leave it out for long analysis.

Opus 5.5 is the first Opus model to launch with Fable-level bio and cyber safeguards

Most flagged messages move to an older model, and the work goes on there. Finding security vulnerabilities in source code is allowed, and 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 and not only from your last message.

In the apps, a notice starting with "Switched to" appears and the chat stays on the older model. To go back, pick Opus 5.5 in the model picker, or start a new chat so the earlier message isn't 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 through /config in Claude Code. Asking the model to reproduce its internal reasoning in the reply is itself a flag category. The guide suggests asking for a short explanation of the chosen approach instead.

The guide admits the safeguards can flag legitimate work, but it gives no rate of incorrect flags. The review comparison rests on a single early tester, and the guide doesn't say how much more fast mode costs. In our view, the sticky switch is the choice to watch. A chat that stays on an older model after a flag can quietly carry a long session there if you miss the notice.

What the flag tuning must fix

Anthropic says it is tuning the safeguards to cut down on incorrect flags, but it gives no timeline or target. Fast mode is still a research preview, and no date has been given for it to leave that status. Until the tuning lands, it's worth setting "Switch models when a message is flagged" the way you want before a long run. Use /feedback whenever a flag looks wrong.