Pawel Huryn · 3 min read · Gen AI Tools
Claude Opus 5: The Best Opus Yet, Once You Heal It
Read the original on productcompass.pm →
The piece traces an initial high regard for Claude Opus 5 followed by growing frustration as real-world use reveals it pausing for permissions and deviating from expected outcomes. It argues these issues stem from a reduced system prompt and RL tuning, and offers a practical fix: codifying three behavioral rules to guide the model’s actions and 'heal' its performance.
Key Takeaways
- Opus 5’s early strengths can be undermined by underlying training tweaks, such as a smaller system prompt and a push toward autonomous tasking.
- Three actionable rules—Act. Don’t Ask; A Question Is a Question; Done Means Done—provide a clear behavioral framework to improve reliability.
- Practical implementation involves adding rule blocks to CLAUDE.md and iterating with concrete commands rather than continuous back-and-forth prompts.
- Applying these rules can materially improve the model’s usefulness, turning Opus 5 into a significantly better tool for the author.
Excerpt
At first I loved Opus 5. On July 24, Anthropic put it at the top of nearly every test they publish: Anthropic's own evaluation table for Claude Opus 5, July 24, 2026. ARC-AGI-3 measures novel problem solving. Opus 5 scored about three times the next best model. But then I used it for a few days. The more I worked with it, the less I liked it . And it’s not just me. For example: My timeline has looked like this for two weeks. “Cancel the subscription.” “Bring back 4.6.” “Anthropic broke it.” The closest thing to advice out there…