Voices
Peter Yang, Shreya Shankar and Hamel Husain publish a free five-step error-discovery skill for Claude Code and Codex
In an August 23 episode, Peter Yang walks through eval design with Shreya and Hamel, splitting evals into top-down rules (their example: 'Is every takeaway 240-330 characters?') and bottom-up failures found by diffing AI output against your own final edits. Their guardrail against overfitting is to compare 10 to 20 past examples before promoting any single failure into an eval. They shipped the five-step error-discovery process as a free skill that runs in Claude Code or Codex; their Maven course claims 4,500+ students and a 4.7 rating across 900 reviews.
↳ Follow the thread