Type: url
Original: https://developers.openai.com/blog/eval-skills/
Status: generated
Summary
A January 2026 post on OpenAI’s developer blog sets out a pattern for testing Codex agent skills with evals rather than judging changes by feel. It recommends defining measurable success first (outcome, process, style and efficiency goals), then running a small set of 10–20 prompts — including negative controls that should not trigger the skill — through codex exec --json, scoring the recorded event trace with deterministic checks, and adding a model-assisted, rubric-based style check constrained by an output schema, growing the prompt set from real failures over time.