A field guide to writing your own eval harness
Why "vibes-based" testing collapses past 50 prompts, and the smallest harness that scales without becoming a second product.
APR 28, 2026 ·
#tooling
Things I have learned about LLMs in 2025
"Annual roundup. Section on tool-calling reliability is gold."
via simonwillison.net APR 16, 2026 ·
The shape of a good MCP server
Most MCP servers I see are CRUD wrappers. Here is what changes when you design tools as if a model were a junior engineer with no memory.
APR 12, 2026 ·
Aider — repo-map heuristics, annotated
"The PageRank-on-symbols trick is more interesting than the headline feature."
via aider.chat MAR 24, 2026 ·