Anthropic just published the prompting guide for Claude Opus 5.5, the model that launched on September 22, and the core message is that teams should stop carrying over their favorite Opus 5 settings without testing them.
The guide tells developers to experiment with multiple effort levels instead of reusing the level they set before. It also suggests that chat applications remove system-prompt lines instructing Claude to think carefully before responding. According to the guide reported by Search Engine Journal, removing that line in a chat product made replies start sooner with “no clear decline in the quality of the reply.”
A new default: medium effort, with thinking always on
Opus 5.5 runs at medium effort by default, one level below the high default on Opus 5. Thinking cannot be switched off in Opus 5.5, and requests that try to disable it will return an error. Anthropic’s guide calls effort the first setting to adjust when balancing quality against speed and cost.
That default shift matters because Anthropic’s testing found Opus 5.5 at medium effort matches or surpasses Opus 5 at high effort on coding and knowledge work. In practice, many teams can get similar output quality at a lower-effort default, then reserve higher levels for harder jobs.
Why the “think carefully” line backfires
On Opus 5.5, the model decides how much thinking a request needs, with effort as the main control. A legacy line that asks Claude to think carefully can therefore add latency without improving the answer. Anthropic’s own documentation still includes such language for Claude Opus 4.7 when effort must stay low, so teams should check whether that line is still appropriate before leaving it in place.
Agent teams should think in time budgets
For multi-agent workflows, the guide recommends giving agents a time budget based on how long a task might reasonably take. Opus 5.5 can track elapsed time, and Anthropic found small agent groups using time signals finish research tasks faster than a single agent working without signals. Those groups still achieved answer quality comparable to solo agents.
The time budget is a suggestion rather than a hard promise, but Anthropic notes a strict timeout can be useful. The tradeoff is that the model may work slightly less thoroughly under pressure.
What to re-check this week
- Test lower effort settings before changing your prompts.
- Audit max_tokens caps carried over from Opus 5, because thinking now uses part of the output cap even when unseen.
- Tag pasted text from emails or documents with a random ID and add a system note on how to handle tagged content.
- Give agent groups explicit time budgets instead of relying on a single agent.
- For frontend projects, specify styles clearly to avoid defaults like cream backgrounds and pill-shaped buttons.
Because the tags are plain text, they provide only one layer of protection against prompt injection.
The bigger shift: controls over coaxing
The guide is part of a wider reset. Several sections ask developers to revisit workarounds inherited from older models, including chart and screenshot instructions and settings written for running with thinking off. Anthropic’s earlier Fable 5.1 guidance made a similar request around formatting rules.
For marketing and growth teams, the takeaway is to treat an AI model migration the way you would treat a platform algorithm change: check your defaults, audit your prompts, and re-test before scaling. An app that never sets an effort value now runs at medium on Opus 5.5 with no code change, which is a sensible starting point, not a permanent configuration.
Source: Search Engine Journal
