Anthropic published this guide to prompting Opus 5.5. Here's the best tips: 1. Do not ask it to "think carefully". There's no quality improvement, and it leads to latency. 2. For long autonomous tasks, Opus may end its work early. It's best to keep track of progress made during the execution or specify what the completed state should look like. 3. For a multiagent setup, give Opus a time budget. Have each subagent report back the `elapsed time / budget`, and it will pace the agents to finish the work in that window. 4. When using data from connected apps like emails and documents, tell it to explore first before acting. 5. For frontend design, clearly state unwanted patterns, not "avoid AI-generated slop". TLDR: ask Opus to "look around" instead of "think carefully".