Skip to main content
EZVals is built for a workflow where you decide what “good” looks like and a coding agent (Claude Code, Codex, Cursor and others) writes the evals, runs them and reads the results. The evals skill teaches your agent how to do that well.

​
Install the skill

ezvals skills add installs the skill version that matches your installed EZVals, so its docs match your code. Choose at least one agent:
  • Combine flags to install for several agents: ezvals skills add --claude --codex. The files are copied once and the other folders link to that copy.
  • The copy goes in .agents/ if you pass --agents, otherwise in the first agent folder in the order above. A local .agents/ folder is added to .git/info/exclude.
  • --global (-g) installs to your home directory, for every project.
After upgrading EZVals, run the same skills add command again to update the skill.

​
What the skill covers

The skill teaches your agent to:
  • plan evals: what to test, how many cases, and where the data comes from
  • wrap your agent as a target
  • choose graders: code checks, LLM judges or human review
  • run evals, read the results, query across runs and suggest fixes
It includes guides for RAG agents, coding agents, agent skills and internal components, plus a copy of these docs matched to your installed version.

​
Prompts to try

Type /evals, or just ask about evals and the agent picks up the skill.
  • “Write evals for the refund flow in my support agent.”
  • “Run the evals and tell me why the failures failed.”
  • “Run each eval 5 times and tell me which ones are unreliable.”
  • “Tighten the LLM judge and regrade the last run.”
  • “Compare the last two runs. What regressed?”
  • “Which run used the most tokens, and on which evals?”
Your agent reads results through ezvals run --json, ezvals query and the run files. You review in the web UI. Annotations and score corrections you make there are saved to the run, so you can ask the agent to read them and adjust the graders.

​
Check or remove the skill

Both accept --global. doctor marks each agent folder as linked, linked elsewhere, a copy, or not installed.