🔬 Cross-Model Benchmarking

Your skill works great in pi. But does it help claude? What about copilot? If you’re building a skill meant to be runtime-agnostic, you need to know it pulls its weight everywhere — including across the models each agent supports.

Cross-model benchmarking runs every eval against every agent you specify — in one command. You get a clean per-model breakdown so you can spot exactly where to focus your tuning.

How it works 🎯

Pass --models with a comma-separated list of agent:model pairs:

skill-eval loop --models pi:claude-sonnet,claude:opus-4-8,copilot:gpt-5

skill-eval runs the full run → grade → benchmark cycle for each model. Each one gets its own with-skill and baseline run, and the results land side-by-side in benchmark.json.

What you get 📊

The benchmark tells you at a glance who’s winning:

{
  "models": {
    "pi-claude-sonnet": { "delta": { "pass_rate": 0.33 } },
    "claude-opus-4-8":  { "delta": { "pass_rate": 0.17 } },
    "copilot-gpt-5":    { "delta": { "pass_rate": 0.33 } }
  },
  "best_model":  "pi-claude-sonnet",
  "worst_model": "claude-opus-4-8"
}

Three signals jump out:

1. best_model / worst_model — no scanning required, skill-eval tells you
2. Per-model delta — is your skill helping pi more than claude? You’ll see it
3. The gap — if one model’s delta is half the others, that’s where to tune your skill next

The iteration loop 🔄

Here’s the real power move. Run with --models across iterations:

Run 1: pi-claude-sonnet +33%,  claude-opus-4-8 +17%,  copilot-gpt-5 +33%
       → Claude's weak at chart-related assertions. Tweak your SKILL.md.

Run 2: pi-claude-sonnet +34%,  claude-opus-4-8 +28%,  copilot-gpt-5 +32%
       → Claude improved! Copilot slipped slightly. Another tweak.

Run 3: pi-claude-sonnet +34%,  claude-opus-4-8 +31%,  copilot-gpt-5 +36%
       → All above 30%. Converging nicely. Ship it! 🚀

Over time, you’ll spot patterns in how different models interpret your skill instructions — and learn to write skills that work well universally.

Composing with --fix 🪄

Cross-model and auto-fix work together seamlessly:

skill-eval loop --models pi:claude-sonnet,claude:opus-4-8 --fix --max-fix-attempts 2

Each model gets its own independent fix phase. If pi passes on the first attempt but claude needs two fix rounds before converging, you’ll see that in the output — and in the benchmark. That tells you claude’s baseline needs more attention than pi’s.

Config for persistence 🛠️

Tired of typing the same --models every time? Set it once in your config:

# .skill-eval.yaml
models:
  - agent: pi
    model: claude-sonnet-4-5
  - agent: claude
    model: opus-4-8
  - agent: copilot
    model: gpt-5

Config models become the default for every run, grade, and loop command. CLI --models always wins when both are present.

⚠️ Cost warning! skill-eval counts your total agent invocations ahead of time. If you’re about to launch 18 runs (3 evals × 3 models × 2 configs), it’ll ask for confirmation first. No surprise bills!

What’s next?

• Auto-Fixing — Combine cross-model with auto-refinement for the ultimate iteration loop.
• Reading Your Results — Make sense of all those deltas and per-eval breakdowns.
• Eval Workflow — The full picture of how everything connects.