Getting the most out of Opus 5.5 in Claude and Claude Code

What it claims

Opus 5.5 behaves differently from Opus 5 in three ways that change how you should prompt it. It runs longer on its own, it always reasons before replying and chooses how much, and it reports plainly what it did. The guide turns these into concrete prompt and CLAUDE.md text. It also documents two non-obvious behaviours:

• Safety switching. Bio and cyber safety checks can silently move a session to an older model, and the session stays on that model.
• Long-chat slowdown. In long chats, the model can re-examine earlier answers, which slows short follow-ups.

Evidence

The page gives anecdotes and internal tests only. It reports no benchmark numbers.

• Prompting for reasoning. In a chat product, removing a "think carefully" line made replies start sooner, with "no clear drop in quality."
• Code review. One early tester said Opus 5.5 at its lowest effort caught more bugs than Opus 5 at high effort, with fewer false alarms.
• Long runs. Early testers ran long coding tasks "for hours with little oversight" and had it coordinate parallel subagents on audits and migrations.
• Detail checking. In internal testing it caught a date that fell on the wrong weekday in a long planning thread, and a chart that contradicted the numbers in a deck.

How to use it

Give it a finish line and an escape condition in one message:

Migrate the payment endpoints from the old client to the new one.
Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes.
Stop and ask me only if a test fails for a reason you can't explain.

Remove "think step by step" style lines from prompts and saved instructions. For a quick reply, say "Answer directly." In Claude Code, adjust how much it thinks with the effort setting instead.

Stop/continue rule for CLAUDE.md. Without this, the model tends to pause with summaries, "want me to continue?" offers, or lists of non-blocking choices.

When a step doesn't need my input, keep going. Put status notes in the same message as your next action.
Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.

Keep permission prompts on for destructive commands as well. For pair programming, ask for the opposite instead: a one-line plan before starting and a recap at the end.

State surviving compaction. "Keep a checklist in TASKS.md. Tick each item when it's done, and add anything new you find." Read that file to track progress, not the scrollback, because Claude Code summarizes older turns as the context fills.

Subagent audits. Give each unit its own subagent. Tell the orchestrator to "check its evidence before you accept it" and to finish with a single table.

Review prompt:

Review the diff on this branch against main.
List only problems you'd block the merge for. For each one, give the file and line, why it's wrong, and how to show it fails.

Design output. The model falls back on default styles, and "avoid a generic look" just swaps one default for another. Name the specific patterns to avoid, for example cream or off-white backgrounds, italic accent words in headings, "01 / 02 / 03" section labels, monospace labels and pill-shaped buttons. Then iterate on whatever it picks instead.

Long-chat slowdown fix (project instructions):

Once you have answered something, treat that answer as done. Focus on what I'm asking now, and don't go back over an earlier answer unless I ask about it or point out a problem with it.

Leave this out of projects doing long analysis, where a later step may expose an earlier error.

Fast mode (/fast in Claude Code) is a research preview. It is the same model with lower latency, but it requires extra usage to be enabled and costs more per token. The guide recommends it for interactive back-and-forth, not for unattended runs.

Limits

• Safety switching. Opus 5.5 ships with "Fable-level" bio and cyber safeguards. Flagged messages move to an older model, shown by a notice starting "Switched to", and the chat stays there.
• The check covers the whole conversation, including files and search results, so a flag can come from earlier content rather than your last message.
• Finding vulnerabilities in source code is explicitly allowed.
• Anthropic acknowledges false positives and says it is tuning the classifier.
• Reasoning requests are a flag category. Asking the model to reproduce its internal reasoning in the reply can be declined or trigger a flag. Remove such requests from prompts and ask for a short rationale instead, e.g. "Explain why you chose this approach in three sentences."
• Removing caution. A keep-going rule reduces stops, so your own safeguards before irreversible actions matter more.
• Thin evidence. Every performance claim is anecdotal or from internal testing, with no stated settings or sample sizes.