Guides · Sep 23, 2026
Anthropic's tips for working with Claude Opus 5.5
Anthropic's own recommendations for Opus 5.5. How to hand over a task, steer long runs in Claude Code and check the result.
Built by Simon + Claude
TL;DR
- The advice is Anthropic's, from the guide Addy Osmani wrote for Opus 5.5. We have picked out what matters most for people who use Claude and Claude Code at work.
- Opus 5.5 works longer on its own, reports more clearly on what it did, and decides for itself how much to think before it answers.
- Their first advice is to give it the whole task in one message, say what done looks like, and delete the lines that tell it to think hard.
This guide is a shorter and hopefully more accessible summary of the recommendations Anthropic's own team has published for Opus 5.5. The advice is theirs, not ours. The original is Getting the most out of Opus 5.5 in Claude and Claude Code written by Addy Osmani and published by Anthropic, with more detail and more example prompts.
This version is written first and foremost for people who use Claude and Claude
Code at work, for analysis, strategy work, proposals and reports, not
developers working in codebases. Many of the tips are for Claude Code, so some
commands, such as /fast, do not work in the regular Claude app. We have added
some visuals and simplified the explanations, and where we add something of our
own, we say so.
If you do not have time to read the whole guide, the shortcut is simple. Start a new session in Claude Code in the project you are working on, share the link to Anthropic's guide, and ask Claude what you should adjust to get the best results with Opus 5.5.
YouRead this guide from Anthropic, https://claude.dev/blog/getting-the-most-out-of-opus-5-5/, and look at my project. What should I adjust to get the best results with Opus 5.5?
Say what done looks like
Anthropic's first piece of advice is to give it the whole task in one message, say what done means, and let it work. Opus 5.5 handles multi-step work better than earlier Opus models, so the most useful thing you can add is the finish line, and the one reason to stop and ask.
A task with a clear finish line is one it can carry to the end without you. In our own guide What is an AI agent this is the loop, where the agent checks the result against the goal and goes round again.
Anthropic also suggests that if you remember something halfway through, you send it as a new message while the task keeps going. Runs are longer now, so starting over costs more.
Delete "think hard"
Anthropic recommends removing lines like "think carefully" from prompts and saved instructions, because Opus 5.5 sets its own amount of thinking on every reply. In their testing, answers started sooner without the lines, with no clear drop in quality. Look for them in project instructions, skills and CLAUDE.md.
Name the styles you do not want
Give Opus 5.5 no direction in design work and it tends to reach for the same few styles. Anthropic's advice is to list the styles you do not want. Their example lists five, and a list like that gives the model more to hold on to than a wish for something modern and clean.
Long runs in Claude Code
Opus 5.5 sometimes pauses to summarise its progress when it could simply have continued. Anthropic suggests a simple rule in CLAUDE.md, the file of instructions Claude Code reads every time it starts. The rule says when to keep working on its own and when to stop and ask you. They also recommend that Claude Code still asks for permission before anything that cannot be undone, such as deleting files. The rule below is a simpler version of their example.
YouWhen a step does not need my input, keep going, and put status notes in the same message as your next action. Stop and ask only when you cannot continue without me, or before anything that cannot be undone, such as deleting files, overwriting other people's work or changing anything outside this folder.
When a job has many parts that can be done separately, Anthropic suggests giving each part to its own subagent, a separate AI agent that takes only that one part. The sentence that matters in their example is the one that asks the main agent to check each subagent's evidence before accepting it. Here it is adapted to a competitor analysis.
YouMap our four closest competitors, with their pricing model, who they sell to and what they say about AI. Give each competitor to its own subagent. Check the sources each subagent brings back before you accept them, and finish with one table of competitor, pricing model, target customers and source.
GritAI comment
Claude Cowork supports subagents too, and it is as simple as asking Claude to use them. Each subagent gets its own context window, so in practice you work with a small team of agents on one task. In return, it uses more tokens.
Long runs fill the context window, which is everything the model can hold in mind at once, and Claude Code then summarises the older parts of the conversation. Anthropic's advice is to keep the checklist in a file such as TASKS.md, which survives the summary.
For back-and-forth sessions where you wait for every reply, Anthropic points to
fast mode, which /fast turns on. It is the same model answering faster, but
the feature is still being tested, costs more, and is paid for with extra usage
(usage credits) outside what your plan covers. On Team and Enterprise, an Owner has to turn
it on for the organisation first.
Checking the result
When a long run is finished, Anthropic suggests reading what it needs from you before the rest of the summary. Their example sets the format once in CLAUDE.md, with three headings at the end of every run.
They also recommend letting Claude review the work before a person does. Their example is about code, where one early tester found that Opus 5.5 on the lowest effort setting caught more bugs than Opus 5 on high effort, and flagged fewer things that were not bugs. We use the same review on proposals and reports.
YouReview the proposal before it goes to the client. List only the problems that are reason enough to stop it going out, and for each one give the page and paragraph, why it is wrong, and what it should say instead.
GritAI comment
This is not new, and it has been part of our training all along. Always ask the AI to quality-check the work first. When the work is complex and involves calculations and analysis, you can also ask the AI to say how confident it is in each fact and calculation, with a confidence interval.
For research, their advice is to ask it to mark what it could not confirm and say where it looked. It costs one sentence and shows you where to check before you pass the answer on.
In the Claude app
Anthropic's guide has four tips for the Claude app, and the first thing to check is that the model picker shows Opus 5.5.
Opus 5.5 has got better with images. Attach the chart or screenshot instead of describing it. According to Anthropic, Opus 5.5 reads images more accurately than Opus 5, such as how the parts of a diagram connect and what changed between two versions.
Ask it to check a long document. In Anthropic's testing, it noticed a date that did not fit its day of the week, and a chart in a deck that disagreed with the numbers.
YouCheck this deck for anything that contradicts itself, in numbers, dates and names. Quote each problem and say where it is.
Ask for the finished file. Ask for the spreadsheet or the document itself, not an outline. According to Anthropic, the spreadsheets and documents it makes need less editing before you share them than Opus 5's did.
GritAI comment
We still recommend a few rounds of brainstorming and gathering context before you make files, because it saves tokens, which is what your usage is measured in. But if you have a clear plan, it is worth testing a delivery end to end, from the request to the finished file.
Say when an answer is done. In long conversations the model sometimes goes back and rethinks questions it has already answered, and then it takes longer to answer what you are asking now. Anthropic suggests an instruction in the project that says what has been answered is done. Leave it out for analysis work, where it helps that the model goes back and fixes an earlier mistake.
YouOnce you have answered something, that answer is done. Answer what I am asking now, and do not go back over earlier answers unless I ask you to or point out that something is wrong.
5 best practices for getting the most out of Opus 5.5
These are the five pieces of advice from Anthropic's guide that we think matter most.
Anthropic's full guide ends with a checklist to run through before your next long task.