Skill Eval Harness

What it does

This skill teaches Claude Code a structured way to test its own work. Before writing code, the agent writes down measurable success criteria (called evals, similar to unit tests for AI tasks), then checks them as it implements and produces a summary report. It also explains how to measure reliability over repeated attempts (pass@k and pass^k) and how to set up regression checks so changes do not break existing behaviour. The guide is written in Spanish, and its examples are shell commands and markdown templates the agent fills in.

Try asking

• “Define pass/fail evals for adding a new user login feature before writing any code”Copy
• “Create a regression eval baseline for our session management and report which checks still pass”Copy
• “Compare how reliable our coding agent is across two model versions using pass@3 and pass^3”Copy
• “Write an eval report for the payment refactor and mark the risky parts for human review”Copy

Get better results

• Write the eval definition before implementation, and keep evals fast so they actually get run
• Use the code-based graders (grep, tests, build) where possible, and reserve model grading for subjective checks
• Review the commands in the examples before running them, since they execute shell commands in your project

What's inside

• SKILL.md