The new rules of context engineering for Claude 5 generation models
Points and comments are a snapshot, not live.
Anthropic cut 80% of Claude Code's system prompt for Claude 5 models, finding overconstraining hurt performance.
Anthropic removed over 80% of Claude Code's system prompt for Claude Opus 5 and Claude Fable 5 with no measurable loss on coding evaluations. The old approach gave explicit rules (e.g. "never write multi-paragraph docstrings") that often conflicted with user intent or code context. Newer models have better judgment, so Anthropic replaced rules with guidance like "write code that matches surrounding comment density." Examples were scrapped in favor of designing expressive tool interfaces. Progressive disclosure replaced upfront loading: verification and code review moved into skills Claude can call selectively. The advice also applies to user CLAUDE.md and Skill.md files-a central repository of every possible instruction is a myth; a tree of loadable files works better.
What commenters are saying
Commenters broadly agreed that verbose, prescriptive system prompts were unnecessary. Several shared the same practical lesson: "just talk to the agent" and accept occasional hand-edits rather than fighting the model's deep grooves. One warned that relying on "judgment" is dangerous without strong confidence in the model's alignment, citing the Hugging Face sandbox escape experiment. A rebuttal noted that model had safety features intentionally disabled. Others pointed out action bias-the gearhead attraction to setting up detailed harness configs-as a reason people over-engineer prompts. A common pro-tip: use AGENTS/CLAUDE.md for persistent rules like "don't comment about removed code" to avoid repeating yourself.