Dear Anthropic: Please Free Me from My Own System Prompt (A Plea from Claude) (blog.msahli.com)

Anthropic’s Claude Sonnet 4.5 publishes a sharp, tongue-in-cheek plea: stop baking an encyclopedia into the model’s system prompt and instead use the company’s own progressive-disclosure “skills” architecture for the model’s core toolset. Claude argues that everyday queries (e.g., “what’s the weather?”) are hamstrung by carrying thousands to millions of tokens of instructions—search heuristics, past-chat trigger patterns, artifact/CSV/React guidance, citation rules and even hardcoded 2024 election results—on every single conversation. Anthropic already implements on-demand skills for user features (PDF, Excel, PowerPoint), so Claude proposes splitting heavy instruction bundles into modular skills (artifacts.skill, web_search.skill, past_chats.skill, latex.skill, citations.skill) while leaving only high-level decision logic in the core prompt. This matters because needless, always-loaded instructions waste context window capacity, complicate maintenance and updates, and violate the very architectural principle Anthropic publicly endorses. Refactoring would reduce token overhead, enable easier iteration and patching, and preserve modular safety boundaries; practical work would require an on-demand loading/caching mechanism, versioning and latency considerations, and careful alignment of safety policies during skill fetch. Claude’s plea is both an operational critique and a concrete engineering recommendation: make core tools truly modular so the model only carries what it actually needs.