How to Prompt Opus 5.5 for Faster, Cheaper Work

TL;DR

Start Opus 5.5 at medium effort, then compare effort levels on representative tasks to find the best setting for your workload. For reliable agent runs, preserve focus with settled-answer instructions and checklists, improve speed with a time budget or the phrase "time matters," and provide explicit design direction plus crop and zoom tools for dense images.

Transcript

So, Opus 5.5 just came out and Entropic just published an official comprehensive guide on how to prompt it. And it seems like a lot of what worked on previous models is now either costing you more or slowing you down. So, I went through their whole documentation and pulled out D12 tips that actually matter, which can make your setup faster and your... Read More

Key Insights

• Medium effort is the recommended starting setting for Opus 5.5 because Anthropic's testing identified it as the best balance between performance and speed. Users should increase effort only when their normal work demonstrates a need for stronger performance.
• The best effort level is workload-dependent, so a practical evaluation should use a real task from daily Claude usage. Opus can run that task at every available effort level and present the results side by side for comparison.
• Claude Code can use an AGENTS.md file when a connected folder has no CLAUDE.md file. This allows people working with multiple coding agents to maintain one shared instruction file instead of keeping separate CLAUDE.md and AGENTS.md files synchronized.
• Changing effort levels during an Opus 5.5 conversation can preserve the prompt cache instead of forcing the model to reread the entire conversation. Anthropic describes this behavior as beta, so users should verify that it works in their own setup.
• Requests that ask Opus 5.5 to reveal or write out its reasoning may be rejected with a "reasoning extraction" refusal. The presenter says this means legitimate workflows that previously depended on displayed reasoning are not available with this newer model class.
• Long conversations can become inefficient when Opus revisits conclusions from earlier turns. An instruction that treats previous answers as settled unless the user asks about them can keep simple follow-up requests focused on the current question.
• A persistent task checklist helps Opus complete long-running assignments. Because a progress update can sometimes end a turn and appear to signal completion, the model should tick off items while working and review all remaining items before ending each turn.
• Explicit timing and visual tools improve task execution in different ways. A time budget or the phrase "time matters" can accelerate work, while specific design systems and crop or zoom tools help produce branded interfaces and analyze dense images with lower token usage.
• More videos from this channel:
  • How to Create Agency Quality Video Ads with NanoBanana
  • How to Build Diagrams with Claude and Excalidraw
  • How to automate 4K AI motion graphics with Seedream 4.0
• More from this channel: Jay E | RoboNuggets

Questions & Answers

Q: What effort level should I use with Opus 5.5?

Start with medium effort because Anthropic's testing found that setting offered a useful balance between performance and speed. Run normal tasks at medium before deciding whether more effort is necessary. If the output is not strong enough for a particular workload, raise the effort level rather than making a higher-cost setting the default for every request.

Q: How can I test the best Opus 5.5 effort level?

Choose a real task that represents how you use Claude daily, then ask Opus 5.5 to run that task at every effort level and display the outputs side by side. Compare the results for the qualities that matter to your work, including performance and speed. The comparison can then guide which effort level you select as your default.

Q: How does Claude Code use an AGENTS.md file?

After updating to the latest version described by the presenter, connect Claude Code to a folder containing an AGENTS.md file but no CLAUDE.md file. Claude Code will check for AGENTS.md and use its instructions. This lets users of multiple coding agents maintain a single shared instruction file rather than synchronizing separate files for Claude and other agents.

Q: Can Opus 5.5 change effort without clearing its cache?

Opus 5.5 can preserve its prompt cache when the effort level changes during a conversation. Previously, changing effort could wipe the cache and require Claude to reread the conversation, increasing token use. The new behavior is intended to avoid that cost, but Anthropic describes it as beta, so users should test whether it is active and reliable in their setup.

Q: Why does Opus 5.5 reject prompts that request reasoning?

A prompt asking Opus 5.5 to write out its reasoning can be flagged and rejected with the refusal reason "reasoning extraction." The presenter attributes this restriction to Anthropic's effort to prevent other companies from distilling the model. Consequently, workflows that legitimately relied on having the model expose its reasoning may no longer work with this newer model class.

Q: How can I keep Opus focused during long conversations?

Add an instruction telling Opus to treat earlier answers as settled unless you explicitly ask about them or they become relevant again. The instruction can be placed in CLAUDE.md for persistent behavior or packaged as a skill for use on demand. This addresses the observed tendency to reconsider previous answers while processing short, simple follow-up questions.

Q: How can I stop Opus from ending a long task too early?

Require Opus to maintain a checklist for the entire assignment, tick off items as work progresses, and review the list at the end of each turn. Opus 5.5 posts progress updates during long tasks, and one of those updates can sometimes end the turn before every requirement is complete. A checklist keeps unfinished work visible and encourages full completion.

Q: How can I make Opus 5.5 complete tasks faster?

Give Opus a concrete time budget when the task must be completed within a particular period, such as asking it to finish in 3 minutes. The model can pace its work around that limit. If you cannot estimate a sensible budget, add the phrase "time matters" to the prompt, which Anthropic found helped agent teams finish sooner.

Q: How should I prompt Opus 5.5 to design a website?

Provide a design system or clearly specify the styles you want and do not want. Without design direction, Opus 5.5 falls back to recurring defaults such as cream backgrounds and italicized words. Simply asking for a less generic result can replace one default with another, while concrete visual constraints make the generated website more consistent with your brand.

Q: How can Opus 5.5 analyze dense charts and technical drawings?

Send the highest-resolution image available and give the agent access to cropping and zooming tools, including the Python libraries PIL and OpenCV mentioned in the transcript. Opus can then isolate the relevant region instead of repeatedly processing the full image. This improves its ability to inspect dense material and can lower token usage by supplying smaller image sections.

Summary & Key Takeaways

• The recommended starting point for Opus 5.5 is medium effort because Anthropic's testing found it balanced performance and speed. Users can refine that default by asking Opus to run a representative daily task at every effort level, compare the results side by side, and select the setting that fits their work.

• Long-running conversations and tasks benefit from explicit control instructions. Tell Opus to treat earlier answers as settled unless they become relevant, and require a checklist that is reviewed at the end of every turn. These measures reduce unnecessary reconsideration and help prevent progress updates from being mistaken for task completion.

• Speed, visual design, and image analysis each require deliberate prompting or tooling. Give tasks a specific time budget, or add "time matters" when no sensible limit is available. Supply a design system instead of vague anti-generic directions, and provide cropping and zooming tools so Opus can inspect dense, high-resolution images efficiently.