Skills for Claude Code and Codex are hard to test. What I mean by hard is that there's no standard way to do it. You evaluate the skill once on something, it looks like it works. You publish it. Then the new super model releases (GLM 5.2 anyone?), it will quietly break for some part, and you won't find out until your users complain.

I also faced the same problem, so I tried to build something lightweight to stop doing that. Caliper.

It's a local and lightweight harness that runs a skill k times in isolated environments and gives you a pass@k score (How much times it succeeded in these k times). As a non-deterministic technology, you can't just say "it worked once". You need to answer how much it passed in k times.

You define success in a YAML spec. I picked YAML to keep a schema and make it still readable for a human. You either use a LLM judge, a Python assertion, or both:

Here's an simple evaluation example with a JSON extraction, so you write this in a YAML file:

  tasks:
    - name: Extracts action items as clean JSON
      prompt: "Read /tmp/transcript.txt and write the
               action items to /tmp/actions.json."
      expect: "A valid JSON array where every item has
               owner, task, due. No markdown fences."
      assert: |
        import json
        items = json.load(open("/tmp/actions.json"))
        assert isinstance(items, list)
        assert all({"owner","task","due"} <= i.keys()
                   for i in items)

Then with the CLI, you'll run it:

caliper run extract-actions.eval.yaml --k 5 --baseline

What's cool about the --baseline flag is that it will re-runs everything without the skill, so you can see whether the skill is doing the work or the base agent was going to pass anyway:

  ID      Task                           k(5)  pass@k
  task-1  Extracts action items as JSON  5/5   100%  PASS
  With skill   100%
  No skill      60%
  Delta        +40%

Most models know how to get the JSON right most of the time (JSON extraction was solved by 2 years old already). But that's it, "most of the time" is the bug. That delta shows how the skill actually helped. (It's sometimes 0%, sometimes -100%!)

I also created two skills you can get started right away with your favorite harness, e.g. Claude Code, Codex or Pi:

- evaluate-skill: run and manage evals without leaving your workflow

- grill-skill: reads your SKILL.md, interviews you about what "good" looks like, writes a 3-task spec (happy path, edge case, adversarial), and runs it

You can install the skill with the command: npx skills@latest add edonadei/caliper

I for now support claude-code, codex, pi, claude-api, openai-api. You can run the agent and the judge as separate backends, so you can run a skill on one and judge with another.

GitHub: https://github.com/edonadei/caliper PyPI: https://pypi.org/project/caliper-eval/

Of course, it's a first step. I think the autorater layer can be vastly improved, more handholding to create and iterate on evaluation specs, supporting more harness, why not including this layer into a self-improvement bigger system?

If you're also building agentic evaluations, I'm genuinely interested to hear how you are handling that.

Customer service as a mirror of modern life