Old Keyword Hinting is expiring, and Claude 5 works better with less intervention

According to Perceive Beating monitoring, Claude Code team member Thariq Shihipar stated that the team has removed over 80% of the system prompts for Opus 5 and Fable 5, with no significant drop in programming assessment.

He mentioned that in the past, to prevent the model from deleting files indiscriminately and writing comments excessively, Claude Code had added numerous prohibitions and examples. The new model is now able to make judgments based on user requests and code style. Having too many rules can actually lead to conflicts.

His suggestion is to streamline CLAUDE.md, only including the project overview and areas where users are most likely to encounter pitfalls. Processes such as testing and code reviews should be broken down into Skills and loaded only when needed. Tools should also guide the model through clear interfaces and parameters, without the need for excessive examples.

Claude Code has also introduced the /doctor command to check whether CLAUDE.md and Skills are too lengthy.