Skills
Stop repeating instructions in the system prompt — put them in tool descriptions and make the tool signature expressive instead
Anthropic's new context-engineering rules invert two long-standing habits: instead of providing tool-usage examples, design expressive tool interfaces and parameter names; instead of repeating instructions throughout the prompt, place each instruction only in the description of the tool it governs. Anthropic attributes its own bloat to conflicting directives fired from multiple layers at once — the example given is a system prompt saying "DO NOT add comments" while a skill says "leave documentation as appropriate." For agent builders this reframes tool schema design as the primary prompt surface.
↳ Follow the thread