Harnessing Claude’s intelligence

Anthropic shares three architectural patterns for building applications on top of Claude that stay effective as the model evolves. First, lean on what Claude already knows: give it general tools like bash and code execution so it can orchestrate its own tool calls and filter outputs without bloating the context window. Second, let Claude manage and persist its own context through skills, context editing, subagents, compaction, and memory folders rather than hard-coding retrieval infrastructure. Third, set boundaries carefully in the agent harness by maximizing prompt cache hits, using declarative tools for security and UX boundaries, and continuously pruning assumptions that newer model versions have made obsolete. Benchmark data from BrowseComp and Pokémon gameplay illustrate how newer Claude versions (Opus 4.5, 4.6) handle memory, context, and long-horizon tasks far better than earlier ones, making previously necessary scaffolding counterproductive.