Lance Martin

Lance Martin works at Anthropic, where he is described as a developer advocate, and the two notes carrying his name both concern the mechanics of running Claude well rather than the models themselves. The February 2026 piece explains prompt caching on the Claude API: cached tokens cost 10% of the normal price, the messages API is stateless so an agent repackages its whole history each turn, and `cache_control` creates a cryptographic hash of every content block up to a breakpoint, scoped to the workspace, with a lookback of at most 20 blocks. The new «auto-caching» mode moves that breakpoint to the last cachable block automatically as the conversation grows. Martin quotes @peakji of Manus calling cache hit rate the most important metric for a production AI agent.

The September 2026 piece, on claude.dev, is about evaluation design and hillclimbing, and adds `/claude-api build-eval` and `/claude-api hillclimb` to the claude-api skill in «Claude Code». His two recommendations are both about not fooling yourself: pick the cases a human can explain the difficulty of, and read a sample of graded transcripts before trusting the grader. Adversarial sampling of cases the current model fails measures that model's failure footprint, not difficulty.

What connects the two is a refusal to treat the harness as free. Caching is a cost line you can read; an eval is a measurement that can quietly lie, and Martin's remedies for both are procedural.