The dial is the model.

{
  "model": "claude-opus-5",
  "max_tokens": 64000,
  "stream": true,
  "output_config": { "effort": "xhigh" },
  "messages": [{ "role": "user", "content": "..." }]
}

At xhigh or max, set max_tokens high — 64k is a sane starting point — because it caps thinking plus response text together. And stream: 64k of output can run past the non-streaming time limit.

Keep responses focused, brief, and concise. Keep disclaimers and caveats short, and spend most of the response on the main answer. When asked to explain something, give a high-level summary unless an in-depth explanation is specifically requested.

Opus 5's default answers run longer than 4.8's. This is the fix — effort is not.

<tone_preference>
Keep outputs reasonably concise.
</tone_preference>

In a long system prompt, pair the full instruction near the top with this short reminder near the end.

Before your first tool call, say in one sentence what you're about to do. While working, give a brief update only when you find something important or change direction. When you finish, lead with the outcome: your first sentence should answer "what happened" or "what did you find," with supporting detail after it for readers who want it.

Opus 5 announces what it's about to do, often at length. Describe the cadence you want; positive examples beat prohibitions.

Match the length of written documents to what the task needs: cover the substance, but do not pad with filler sections, redundant summaries, or boilerplate.

Files written to disk — reports, Markdown, summaries — bloat independently of conversational verbosity. This is a separate lever.

Deliver what was asked, at the scope intended. Make routine judgment calls yourself, and check in only when different readings of the request would lead to materially different work. If the request seems mistaken or a better approach exists, say so in a sentence and continue with the task as asked rather than quietly narrowing, widening, or transforming it. Finish the whole task, and stop short of actions that are clearly beyond what was asked.

The best all-purpose block on this page. Opus 5 will apply its own judgment about what the task "should" be unless you draw the line.

Delegate to a subagent only for large tasks that are genuinely independent and parallelizable, such as a wide multi-file investigation. Do not delegate work you can finish yourself in a handful of tool calls, and do not use subagents to verify or double-check your own work. If one subagent can complete the task, use one rather than several, and keep spawn counts low.

Delegation multiplies cost and wall-clock on small tasks. Opus 5 reaches for it more readily than 4.8 did.

Only correct an earlier statement when the error would change the user's code, conclusions, or decisions. State corrections plainly and briefly, then continue the task. For slips that change nothing for the user, make the fix and move on without noting it.

Self-correction is a feature; narrating every self-correction to a customer is not.

Review this change and report every issue you find, including low-severity ones and anything you are only moderately confident about. Do not filter for severity — I will triage in a separate pass. For each finding give the file and line, a one-sentence statement of the defect, and a concrete failure scenario with inputs and the wrong result.

Precision and recall are both high on Opus 5, and accuracy holds at lower effort — so run a fast medium pass at review time and a thorough xhigh pass later.

You may say a brief sentence before using a tool.
Do not include internal or system XML tags in your response.

With thinking off, Opus 5 occasionally writes a tool call as plain text (it never runs, and it poisons later turns) or leaks tags. These two lines mitigate both. Better: keep thinking on and drop effort instead.