How to Get the Most Out of Claude Opus 5.5

Anthropic published a playbook on working with Claude Opus 5.5, the flagship model that is also available on VEGA AI. The short version: Opus 5.5 works on its own for longer than Opus 5 did, it says plainly what it did, and it thinks before every reply. Below is the whole guide, with the text from the article's figures transcribed as code so you can copy it.

Three things to try first

• Hand over the whole task in one message: what "done" looks like, and when you want it to stop and ask. Then let it work.
• Delete the "think carefully" lines from your prompts and saved instructions. Opus 5.5 already thinks before every reply.
• When a long run ends, read the part that needs you first, before the rest of the summary.

How to ask

Give it a finish line, then let it run

Opus 5.5 holds up on long, multi-part work much better than Opus 5 did. Early testers had it run coding tasks for hours with little oversight, and its biggest gains over prior Opus models are exactly there: carrying a change through a large repository until the tests pass. For that to work, it needs to know where the finish line is.

The article shows the recommended prompt shape as a card with three labeled boxes. Transcribed:

THE WHOLE TASK
  Migrate the payment endpoints from the old client to the new one.

THE FINISH LINE
  Done means: every endpoint uses the new client, the old client
  is deleted, and the test suite passes.

WHEN TO STOP
  Stop and ask me only if a test fails for a reason you can't explain.

The card's footer sums it up: give the whole task in one message, name the finish line, then leave it alone.

Stop telling it to think hard

Remove "think carefully" and "think step by step" from your prompts and saved instructions. Opus 5.5 always thinks before it replies and decides how much on its own. In Anthropic's testing in a chat product, deleting a "think carefully" line made replies start sooner with no clear drop in quality. For a quick answer to a simple question, say "Answer directly". In Claude Code, the effort setting controls how much it thinks.

Add to a running task

If you remember something mid-run, type the follow-up while it works. In Claude Code you write the message and press Enter, for example "Also keep the old endpoint names as aliases." This matters more than before because runs are longer, so a restart costs more.

For design work, name the styles you don't want

With no design direction, Opus 5.5 falls back on a few default styles, and a vague "avoid a generic look" mostly swaps one default for another. A list of specific patterns works much better:

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.

If you don't like what it chose instead, add that to the list and ask again.

Steering a long run in Claude Code

Tell it which stops you want

Opus 5.5 keeps you posted as it works, and on a long task it sometimes stops to report instead of going on: a summary that names the next step without taking it, an offer to continue, or a list of choices that don't block the work. The article suggests a CLAUDE.md rule, shown in the figure as two boxes. Transcribed:

KEEP GOING
When a step doesn't need my input, keep going. Put status notes in the
same message as your next action.

STOP AND ASK
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.

If a run still stops with "Want me to continue?", reply "continue". A keep-going rule means fewer stops, so the last line matters: permission prompts for destructive commands should stay on. For pair programming you may want the opposite rule, a one-line plan before it starts and a short recap at the end. Opus 5.5 follows either one.

Split big work across subagents

For an audit, a migration, or a review across a large codebase, ask Opus 5.5 to fan the work out to subagents and check each result. Early testers had it coordinate parallel subagents on long audits and migrations with little oversight. The figure in the article is a flow diagram; in text form:

Audit every service in services/ for the retry bug in the linked issue.
                          |
        +---------+---------+---------+
        v         v         v         v
   [Subagent] [Subagent] [Subagent] [Subagent]
        |         |         |         |
        +---------+---------+---------+
                          v
                  CHECK ITS EVIDENCE
   When a subagent reports back, check its evidence
   before you accept it.
                          |
                          v
               Finish with one table.

| SERVICE | AFFECTED YES OR NO | THE EVIDENCE |
|---------|--------------------|--------------|

And the prompt itself:

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

Long runs fill the context window, and Claude Code then summarizes older turns. A checklist in a file survives that, and it shows what is done and what is left at a glance: "Keep a checklist in TASKS.md. Tick each item when it's done, and add anything new you find." Read the file, not the scrollback, to see where the run is.

Checking the result

When a long run ends, look first for anything Claude is waiting on you for, like a decision it left open or a change it wants you to approve. Then read the rest of the summary. To change the format, say so in CLAUDE.md: "End every run with three headings: Blocked on me, Changed, Found."

You can also ask Opus 5.5 to review a diff or a pull request before a person does. One early tester said it caught more bugs at its lowest effort than Opus 5 at high effort, with fewer false alarms:

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.

For research and analysis, add "Mark anything you couldn't confirm, and say where you looked" to the request. It works in a Claude research report and in Claude Code, and it makes the gaps easy to spot.

In the Claude apps

First check that the model picker says Opus 5.5. Then:

• Attach the chart, diagram, screenshot, or slide itself instead of retyping the numbers. Opus 5.5 reads images more accurately than Opus 5, including meaning that depends on where things are: which boxes an arrow connects, what changed between two versions of a diagram, when a meeting starts and ends in a calendar screenshot. Ask a specific question, like "Which of these services call the billing API directly?"


• Give it a long plan, report, or deck and ask it to find mistakes: "Check this deck for anything that contradicts itself: numbers, dates and names. Quote each problem and say where it is." In Anthropic's testing it caught a date that fell on the wrong weekday and a chart that didn't match the numbers in the deck.


• When you want a spreadsheet or a document, ask for the finished file rather than an outline: "Make this a spreadsheet I can share: one row per vendor, with columns for cost, contract end date and owner."


• If follow-up questions in a long project chat feel slow, the model may be re-checking earlier answers. Add this to the 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 it out of projects for long analysis, where a later step can show a mistake in an earlier one.

When a message is flagged

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 education questions still work. The safeguards can flag legitimate work though, and Anthropic says it is tuning them to cut incorrect flags.

In the Claude apps, a flagged message produces a "Switched to" notice naming the older model, and the chat stays on it. To go back, pick Opus 5.5 in the model picker; if the earlier message is still in the chat it may be flagged again, so starting a new chat is cleaner. To be asked first instead, turn off "Switch models when a message is flagged" in Settings, then Capabilities, and you will get a paused card with options. The check covers everything in the conversation, including files and search results, so a flag can come from earlier content.

In Claude Code, the session continues on the older model. Run /model to switch back, press Esc twice to edit your last message and retry, change the same setting under /config to be asked first, and run /feedback if the flag was wrong.

One related prompt edit: asking the model to reproduce its internal reasoning in the reply can be declined, because it is one of the flag categories. Ask for what you actually need instead, for example "Explain why you chose this approach in three sentences."

Fast mode

Fast mode is available for Opus 5.5 at launch as a research preview in Claude Code: same model, text arrives sooner, costs more per token, and needs extra usage turned on. Type /fast. It suits back-and-forth work where you read each reply before sending the next.

The checklist

The article ends with a checklist card to run through before a long task. Transcribed:

ASKING
[ ] The task says what "done" looks like
[ ] No "think hard" lines in prompts or saved instructions
[ ] Design requests list the styles to leave out
[ ] Charts and screenshots are attached, not retyped

LONG RUNS IN CLAUDE CODE
[ ] CLAUDE.md says when to stop and when to keep going, and to stop
    before anything destructive
[ ] Permission prompts are still on for destructive commands
[ ] Large audits and migrations are split across subagents
[ ] The task list is kept in a file

CHECKING
[ ] The "needs from you" part of the report is read first
[ ] A review pass runs before a person reviews
[ ] Research answers mark what couldn't be confirmed

FLAGS
[ ] You know how to switch back: the model picker, or /model
[ ] "Switch models when a message is flagged" is set the way you want

Claude Opus 5.5 is available on VEGA AI in the model picker.