# Claude Code cut its system prompt 80%: trim your CLAUDE.md

> Claude Code's system prompt was reportedly cut 80%. With stronger models, lean context beats long rule lists, so trim CLAUDE.md and move detail into skills.

Published: 2026-07-26
URL: https://daniliants.com/insights/claude-code-cut-its-system-prompt-80-trim-your-claude-md/
Tags: claude-code, context-engineering, system-prompts, skills

---

## Summary

The creator claims Anthropic cut the Claude Code system prompt by 80%, and argues that with newer, stronger models, shorter and less prescriptive context often outperforms long rule lists, because the model gets to decide what context it needs rather than being told upfront. The same logic applies to a project's CLAUDE.md: trim it down and move detail into skills the agent can pull in only when needed (progressive disclosure).

## Key Insight

- Core claim: newer and stronger models (the video cites Sonnet reaching Opus-level performance in the same harness purely via context optimization) do better with lean, permissive context than with long prescriptive rule lists, and over-specifying can hurt rather than help.
- The mechanism argued: giving a model fewer upfront rules lets it make its own judgment calls about what to load and when, instead of being boxed in by exhaustive instructions.
- The same logic is applied to CLAUDE.md files, both user-level and project-level, as to system prompts. The claim is that most CLAUDE.md files are bloated for what current models actually need.
- Recommended pattern: move detail out of the always-loaded CLAUDE.md into skills, so the agent progressively discloses information only when a task actually needs it, rather than paying the context-window tax on every turn.
- Caveat: the video is thin on specifics. No named papers, no numbers on how much to cut or which sections tend to be dead weight, and the reference to "dozens of independent research papers" is asserted without citation. Treat it as a directional signal, not a validated methodology.