Anthropic Reduces Claude Code System Prompt by 80%

Anthropic revealed that it reduced Claude Code's system prompt by 80% for Fable 5 and Opus 5 models without any loss in coding benchmarks. In a fireside chat, product lead Cat Wu and engineer Thariq Shihipar explained that removing few-shot examples improved creativity, and replacing hard prohibitions with contextual guidance reduced conflicts. The team now scrutinizes every instruction for 100% accuracy, softening or removing those with exceptions. This approach relies on the advanced judgment of frontier models and would not work for smaller models like Haiku. Key trends include shifting from rigid rules to judgment-based prompts, from upfront context to progressive disclosure via Skills, and from manual memory to automatic memory. The team also shared that 65% of Anthropic's product engineering PRs are now handled by Claude Tag, a collaborative Slack agent. The findings suggest a paradigm shift in prompt engineering as models become more capable.
Key insights
Prompt engineering is evolving from exhaustive instruction sets to minimal, judgment-based guidance as models improve
Anthropic found that removing few-shot examples made Claude more creative, and that hard prohibitions often conflicted with user instructions. The new approach gives models more context and trusts their judgment, with the rule that every instruction must be 100% correct in all scenarios. This represents a fundamental shift from the long-held best practice of providing detailed examples and constraints.
The 80% reduction demonstrates that over-constraining models can be counterproductive, causing conflicts and limiting creativity
The team discovered that many previous instructions were only 90% correct—such as mandatory frontend verification after any change, which is unnecessary for trivial string edits. By softening such rules and letting the model use its own judgment, performance was maintained while reducing prompt size dramatically.
This approach is model-dependent; smaller models still require detailed prompts, creating a bifurcation in best practices
Simon Willison noted that this method only works for Opus/Fable-level models; Haiku and similar small models still need detailed prompts. Anthropic maintains separate system prompts per model, acknowledging that frontier models can handle ambiguity that weaker models cannot.
The shift toward progressive disclosure and modular Skills reflects a broader trend of treating context as a dynamic, on-demand resource
Instead of loading all context upfront, Anthropic now separates information into Skills that are called on demand, uses automatic memory instead of manual claude.md entries, and prioritizes code references over descriptions. This reduces cognitive load on the model and improves scalability.
Anthropic's internal adoption of Claude Tag for 65% of PRs signals a growing reliance on AI agents for core engineering work
Claude Tag, a Slack-integrated collaborative agent, now handles 65% of product engineering PRs at Anthropic. Complex tasks use Claude Code interactively, while routine work is delegated to Claude Tag. Trust was built over six months by adding every incident-causing PR to the evaluation set.