How to Prompt Claude Opus 5 for Better Results

TL;DR

Give Opus 5 the complete job upfront, define where the work ends, and constrain both its final response and the size of the deliverable. Start with low or medium effort for ordinary knowledge work, and avoid demanding redundant post-completion checks because the model already reviews and fixes its work as it proceeds.

Transcript

Claude Opus 5 is here, and on paper, it is the best Opus model Anthropic has ever released, sometimes even better than Fable 5. But if you prompt it the same way you've been prompting previous Claude models, you are not going to get good results with it. Anthropic agrees with this, so they published an entire prompting guide for Opus 5. It is dense... Read More

Key Insights

• Opus 5 works best when the complete assignment is provided at the start, including the available inputs, the intended outcome, and where the finished work should go. Breaking a connected project into many small preliminary requests may be unnecessary for this model.
• Model selection depends on the task and subscription plan. The suggested approach is to use Sonnet as the default for knowledge work on the $20 plan, reserve Opus 5 for harder tasks, and consider Opus 5 as a regular default on Max plans.
• Effort settings mainly control how much thinking time and how many tokens Claude uses. Higher effort does not automatically guarantee a better result, and Anthropic describes low and medium effort as surprisingly strong for many ordinary knowledge-work assignments.
• Clear stopping conditions are important because Opus 5 may complete more work than requested. A prompt should define the exact deliverable, specify what must remain unchanged, and state when questions are justified so the model has freedom without an open-ended mandate.
• Response-length limits control Claude's explanation after completing the task. Users can request a fixed number of short bullets, the location of the finished work, and only essential operational notes to avoid receiving an unnecessarily long account of the process.
• Deliverable-size limits are separate from response-length limits. A short completion message does not prevent an oversized website or document, so the prompt should also constrain sections, word counts, headline length, reading time, or another measure appropriate to the output.
• Redundant checking instructions may consume extra time and tokens without improving the result. Anthropic says Opus 5 checks and fixes its work while proceeding, so users can still define quality standards without ordering repeated reviews of every file after completion.
• The Northline Coffee test showed that precise scope can reduce follow-up questions and accelerate execution. Claude Design built the requested six-section website in less than three minutes, incorporated the supplied assets, and returned exactly the requested three explanatory bullets.

Questions & Answers

Q: How should I structure a prompt for Claude Opus 5?

Start by giving Opus 5 the entire job, including the source materials, objective, desired final result, and destination for the completed work. Then establish boundaries covering what it should build, what it must preserve, and when it may ask questions. Finally, constrain both the size of the deliverable and the length of the completion message.

Q: Which Claude model should I use for knowledge work?

The suggested choice depends on the task and plan. On the $20 Claude plan, use Sonnet as the default for knowledge work and move to Opus 5 for more complex assignments or unsatisfactory Sonnet results. On a Max plan costing $100 or more, Opus 5 can often serve as the default for most knowledge-work tasks.

Q: When should I switch from Opus 5 to Fable 5?

Switch to Fable 5 when Opus 5 produces a result that does not satisfy you and you want to compare the two models' outputs. Although Fable 5 is described as the most powerful and most expensive option, Opus 5 beats it on many benchmarks according to the transcript, making Fable less necessary for routine use.

Q: What effort level should I choose for Opus 5?

Start with low or medium effort for normal knowledge work because the setting mainly affects how much thinking time and how many tokens Claude spends. The suggested simple defaults are low effort for users on the Pro plan and medium effort for users on Max plans. Higher effort can help sometimes, but it may also be unnecessary overkill.

Q: Why should Opus 5 receive the full job upfront?

Opus 5 is presented as capable of taking a complete assignment and working through its connected stages independently. Instead of separately requesting brand extraction, a brand guide, a website plan, and implementation, describe the starting assets and finished outcome together. This gives the model a clear path from the available materials to the requested deliverable.

Q: How do I stop Opus 5 from doing too much work?

Define exactly where the assignment ends and list the important limits. For the coffee website, the prompt requested only a home page, preserved the existing logo, colors, and product names, and connected buy buttons to an existing shop page. It also allowed small design decisions while reserving questions for issues that would change the whole website.

Q: How can I make Opus 5 give a shorter final response?

State precisely what the completion message should contain and how long it may be. The example asked Claude to identify where the website was located, summarize the work in no more than three short bullets, and mention only information that was truly important for understanding how the website worked. Claude returned exactly three explanatory bullets.

Q: Should I ask Opus 5 to check all of its work twice?

Do not automatically require Opus 5 to repeat a complete review after finishing. Anthropic says the model already checks and fixes its work as it proceeds, so another mandatory inspection may increase time and token use without improving the final result. Continue describing what good work looks like, but omit repetitive checking instructions unless the task specifically requires them.

Summary & Key Takeaways

• Opus 5 performs best when it receives the full assignment at the beginning instead of a sequence of narrowly separated steps. A strong prompt identifies the starting materials, desired outcome, project location, and business goal, giving the model enough context to plan and complete the entire workflow independently.

• Clear stopping conditions prevent Opus 5 from expanding a task beyond what the user actually needs. Prompts should identify what to build, what to preserve, which existing resources to use, how much independent discretion is allowed, and the limited circumstances under which the model should interrupt the work with a question.

• Users should separately control the deliverable and the completion message. The Northline Coffee example limited the website to six scannable sections and restricted the final report to three short bullets. With medium effort and detailed boundaries, Claude Design produced the requested structure in less than three minutes without follow-up questions.