Anthropic just deleted 80% of Claude Code's system prompt
Anthropic published new guidance on July 24 for building with the Claude 5 generation, Opus 5 and Fable 5, and the headline number is blunt: they cut more than 80% of Claude Code's system prompt and lost nothing. All those carefully worded rules and examples that used to guard the model turned out to be dead weight.
The old playbook was give the model explicit rules, front-load every scrap of context, repeat your instructions, stuff memory into CLAUDE.md. The new one flips all of it. Let the model use its own judgment. Design better tools instead of writing usage examples. Reveal context progressively instead of dumping it up front. Let memory manage itself. Point the model at rich references, actual code, artifacts, rubrics, instead of hand-holding markdown.
Here's why it matters. For two years, prompt engineering was really about not trusting the model, building fences so it wouldn't wander off. Anthropic is now saying the fences cost more than they save. A model that can handle ambiguity doesn't need you narrating every step, and every token you spend narrating is a token it can't spend thinking.
If you build agents, treat this as a rewrite of your priors. Stop engineering the prompt, start engineering the tools and the environment the agent lives in. Full post: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
← Back to all articles
The old playbook was give the model explicit rules, front-load every scrap of context, repeat your instructions, stuff memory into CLAUDE.md. The new one flips all of it. Let the model use its own judgment. Design better tools instead of writing usage examples. Reveal context progressively instead of dumping it up front. Let memory manage itself. Point the model at rich references, actual code, artifacts, rubrics, instead of hand-holding markdown.
Here's why it matters. For two years, prompt engineering was really about not trusting the model, building fences so it wouldn't wander off. Anthropic is now saying the fences cost more than they save. A model that can handle ambiguity doesn't need you narrating every step, and every token you spend narrating is a token it can't spend thinking.
If you build agents, treat this as a rewrite of your priors. Stop engineering the prompt, start engineering the tools and the environment the agent lives in. Full post: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
Comments