Claude Code cuts 80% of its system prompt and shifts to on-demand Skills

APPSO · wechat · 2026-07-27

Claude Code trims 80% of its system prompt as models get smarter

The article argues that Anthropic’s Claude Code team is moving from heavy prompt-and-rule engineering toward shorter instructions, better tool interfaces, and on-demand context loading.

The team reportedly removed more than 80% of Claude Code’s system prompt for newer models such as Claude Opus 5 and Claude Fable 5, without seeing an internal coding benchmark drop.
Many old prompt rules existed mainly to compensate for weaker models; with better models, they now create context bloat and even instruction conflicts between system prompts, Skills, and user instructions.
The new approach emphasizes:
Trusting the model’s judgment instead of hard-coding every rule.
Designing clearer tools so the model can infer the right action from interface states.
Progressive disclosure: load Skills, code-review rules, verification steps, or tool docs only when needed.
Removing duplicated instructions and letting the model retain useful working memory automatically.
Using richer references such as HTML prototypes, tests, code, and scorecards instead of only Markdown specs.

The post reframes prompt writing as environment design: for stronger agentic models, humans should focus less on micromanaging behavior and more on defining good interfaces and good outcomes.

Original post →

Claude Code cuts 80% of its system prompt and shifts to on-demand Skills

More from coding & agent