Thariq Shihipar: Old prompt-wording experience no longer works; the less Claude 5 is restricted, the better it performs

robot
Abstract generation in progress
Crypto news: Claude code team member Thariq Shihipar posted that the team removed more than 80% of system prompt words for Opus 5 and Fable 5, and that programming evaluations have not shown a noticeable decline. He said that in the past, to prevent the model from randomly deleting files and abusing comments, Claude code added a large number of prohibitions and examples. The new model can already make its own judgment by combining user requirements and coding style; writing too many rules risks them conflicting with each other. He suggested simplifying claude.md to only include the project introduction and the parts that are genuinely easy to trip over, splitting complex processes like testing and code review into skills and loading them only when needed. The tool should guide the model through clear interfaces and parameters, without piling on examples. Claude code also added a /doctor command to check whether claude.md and skills are too long.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned