Claude Opus Best Practices

Anthropic publishes a short playbook with each new Opus model. This note keeps them together, newest first, so I can see what changed between versions and which habits to drop.

Opus 5.5

23 Sep 2026

Source: Getting the most out of Opus 5.5 in Claude and Claude Code, by Addy Osmani on claude.dev, 22nd September 2026.

The playbook names 3 behaviours that differ from Opus 5:

• It works longer on its own. Its biggest gains are on multi-step work, like carrying a change through a large repository until the tests pass. Early testers ran coding tasks for hours with little oversight.
• It reports plainly. Updates and final summaries say what it did, what it found, and what it needs from you.
• It always thinks before it replies, and decides how much by itself.

The 3 things to try first: hand over the whole task with a clear finish line, delete every "think carefully" line, and read what it needs from you before anything else when a run ends.

How to ask

Say what "done" looks like, then let it run. Give the whole task in 1 message, with the finish line and the condition for stopping:

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.

Stop telling it to think hard. Remove "think carefully" and "think step by step" from prompts and saved instructions. In Anthropic's testing, removing the line made replies start sooner with no clear drop in quality. For a quick answer, say "Answer directly." To change the depth in Claude Code, change the effort level instead.

Add to a running task. Type a follow-up and press Enter while Claude works. Runs are longer now, so a restart costs more.

For design work, list the styles you don't want. Without direction, Opus 5.5 falls back on a few default looks. "Avoid a generic look" only swaps one default for another. A list of specific patterns works:

Build a personal website with placeholder content.
Don't use a cream or off-white background, italic accent words in
headings, numbered "01 / 02 / 03" section labels, monospace labels, or
pill-shaped buttons.

Then look at what it picked instead, and add that to the list if you don't like it either.

Steering a long run in Claude Code

Write down which stops you want. On a long task, Opus 5.5 sometimes stops to report instead of going on: a summary that names the next step without taking it, or "Want me to continue?". It follows a rule in CLAUDE.md that names the stops:

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.

Fewer stops means fewer chances to catch a mistake, so keep the last line, and keep permission prompts on for destructive commands. For pair programming, write the opposite rule: a 1-line plan before it starts, and a short recap at the end.

Split big work across subagents, and check their evidence. For audits, migrations and reviews across a large codebase:

Audit every service in services/ for the retry bug in the linked issue.
Give each service to its own subagent. When a subagent reports back,
check its evidence before you accept it.
Finish with one table: service, affected yes or no, and the evidence.

Keep the task list in a file. A long run fills the context window, and Claude Code then summarises older turns. A checklist in a file, for example TASKS.md, survives that, and shows at a glance what's done and what's left.

Checking the result

• Read what it needs from you first: open decisions and changes waiting for approval. Then the rest of the summary. To set the format, add to CLAUDE.md: "End every run with three headings: Blocked on me, Changed, Found."
• Let it review 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. Ask for merge blockers only, each with the file, the line, why it's wrong and how to show it fails.
• Ask it to mark what it couldn't confirm, and where it looked. "I couldn't find this" is worth reading.

In the Claude apps

• Attach the chart or screenshot instead of retyping the numbers. Opus 5.5 reads images more accurately than Opus 5, including meaning that depends on position: which boxes an arrow connects, or what changed between 2 versions of a diagram.
• Ask it to check a long document for contradictions in numbers, dates and names. In Anthropic's testing it caught a date on the wrong weekday and a chart that didn't match the numbers in a deck.
• Ask for the finished file, not an outline. Its spreadsheets and documents need less editing than Opus 5's.
• In a long project chat, say when answers are settled. Opus 5.5 sometimes goes back over an earlier answer while it thinks about a short follow-up, which slows the reply. A project instruction fixes it: "Once you have answered something, treat that answer as done." Leave it out of projects for long analysis, where a later step can expose an earlier mistake.

When a message is flagged

Opus 5.5 is the first Opus model to launch with the same bio and cyber safeguards as Fable. 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 education questions should still work. Anthropic says the safeguards sometimes flag legitimate work and are being tuned.

The check covers the whole conversation, including files and search results, so a flag can come from earlier content. And don't ask it to reproduce its internal reasoning in the reply: that request is one of the flag categories. Ask for a short explanation of the chosen approach instead.

Speed

Fast mode (/fast in Claude Code) serves the same model with faster output. At launch it's a research preview, needs extra usage turned on, and costs more per token. Use it for back-and-forth work where you read every reply.

What changed since Opus 4.7

The April advice further down still mostly holds. 1 point is now reversed.

What it means for my setup

• No "think hard" lines to delete. I searched my instruction files on 23rd September 2026 and found none.
• The task list in a file is already how I work. My TASKS.md queue does this, as described in The NicAI workspace.
• The stop rule is worth adding to my global CLAUDE.md, including the line about destructive actions.
• The end-of-run format fits next to mine. My runs already end with a list of deliverables and a stats line (see My Claude Code setup ). "Blocked on me" at the top would make long runs faster to review.

Opus 4.7

17 Apr 2026

Task spec and structure

Front-load the task. Instead of revealing requirements progressively across turns, hand over intent, constraints, acceptance criteria, and file locations in the first prompt. Batch clarifying questions. Every turn adds reasoning overhead. For long-running safe work, use auto mode (Shift+Tab) so the model can iterate without frequent check-ins.

Effort levels

xhigh is the recommended default for coding and agentic work.

Adaptive thinking

Extended thinking with fixed budgets is gone. Opus 4.7 decides when thinking is worth it. Influence behavior with natural language:

• More thinking: "Think carefully and step-by-step before responding."
• Less thinking: "Prioritize responding quickly rather than thinking deeply."

Behavioral changes

Response length now matches task complexity. State explicit style preferences in the prompt if you want something different. The model reasons more and uses tools less frequently. Subagent spawning is more judicious. For fanned-out work, explicitly ask for parallel agents.

Takeaways

1. Stop dribbling context. Specs go in the first message.
2. xhigh is the default. Don't bump to max out of habit.
3. If parallelism matters, ask for it by name.
4. Control thinking depth with plain English, not dials.

Further Reading

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

claude.dev playbook, 22nd September 2026

https://claude.dev/blog/getting-the-most-out-of-opus-5-5/

Best practices for Claude Opus 4.7 with Claude Code

Anthropic blog, official guidance

https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code