Anthropic Shares Three Patterns for Building With Claude

Anthropic published an architectural guide for developers building on Claude, organized around three patterns: leverage what Claude already knows (49% on SWE-bench Verified using just bash and text editor), systematically remove scaffolding as capabilities improve, and set boundaries through declarative tools rather than restrictive prompts. The standout data point - letting Opus 4.6 filter its own tool outputs on BrowseComp brought accuracy from 45.3% to 61.6%. The guide's real message is that most teams are over-engineering their Claude integrations. Pre-loading system prompts, routing every tool result through the context window, and building elaborate retrieval systems are patterns that made sense with weaker models but now actively limit performance. OpenAI and Google have published similar "less scaffolding" guidance for their APIs this quarter, and the convergence is telling.