Anthropic Just Revealed How to Prompt Opus 5

Key technological/product ideas

• Claude Opus 5 is Anthropic’s flagship “best Opus” (sometimes claimed to be even stronger than Fable 5 on benchmarks), but it won’t respond well if you prompt it the same way as older Claude models.
• Anthropic provides a dedicated prompting guide for Opus 5, implying prompting strategy is a major factor in getting high-quality outputs.

Model choice + effort level (workflow guidance)

Available models (4 total)

• Haiku = weakest (generally not recommended for serious work)
• Sonnet = recommended default for many users on lower-tier plans
• Opus 5 = often preferred default on higher tiers
• Fable 5 = most powerful/expensive; can be useful for comparison, but “less necessary” with Opus 5

Effort level isn’t “more = better”

• Effort mainly controls how much “thinking” time/tokens the model spends.
• Low and medium effort are surprisingly strong on Opus 5 for normal knowledge work.

Suggested defaults:

• Low for users on Pro
• Medium for users on Max plans
• For the creator: Opus 5 + medium is the default.

The “5 prompting rules” for Opus 5 (tutorial-style)

Using a step-by-step example prompt for a company (Northline Coffee), the video extracts five practical rules for better results:

1. Give the full job up front (don’t over-break into steps). Instead of “extract branding → create guide → plan website,” provide a complete spec like: build the website end-to-end using provided assets and output into the project folder.

2. Define where the job ends + add clear limits. This prevents Opus 5 from doing “extra work” beyond what’s needed. Example limits:

   • build only the home page
   • link buy buttons to an existing shop page
   • keep brand assets unchanged
   • ask questions only if it would change the whole site

3. Constrain output length for the text response. Opus 5 may otherwise produce overly long summaries/instructions. Example: after completion, provide where the website is and no more than three short bullets, plus only essential technical/operational notes.

4. Constrain the size/structure of the deliverable (not just text). Even if the response text is short, the generated artifact can be too large. Example constraints:

   • keep the website to six sections
   • apply tight structure rules like:
     • headline ≤ 8 words
     • each section ≤ ~70 words
     • easy-to-scan structure

5. Don’t ask Opus 5 to repeatedly “check every file twice” / do redundant self-audits. Anthropic claims Opus 5 already checks and fixes while generating. Forcing full re-checking wastes time/tokens without improving results.

Tool usage mentioned (product workflow)

The tutorial applies whether you use Opus 5 via:

• Claude Chat
• Claude Co-work
• Claude Design
• Claude desktop/app and other integrations

The video demonstrates using Claude Design to generate a website.

Practical results from the demonstration (review/analysis)

• The example website generation completed in under ~3 minutes (noting it’s usually slower for the creator).
• The output:
  • adhered to the requested six sections
  • followed the “3 short bullets only” instruction in chat
  • did not ask follow-up questions, attributed to the prompt containing the needed details

Main speakers/sources

• Primary speaker: The YouTube creator/host (no specific name provided in the subtitles)
• Primary source referenced: Anthropic (Opus 5 prompting guide + statements about effort levels and built-in checking behavior)