Home

Skill Harness

An evaluation harness for reusable AI agent skills — a skill disposition engine. It measures whether a skill earns its slot by running the same task with it and without it, and comparing. First-class Claude Code support, built to extend to other agent ecosystems.

You get one of three honest answers:

• KEEP — the skill measurably earns its slot.
• CUT — either subsumed (the model already does the task without the skill) or no-lift (it was measured on a task the model needed help with, and it didn't deliver).
• CAN'T-TELL-YET — the evidence doesn't support a verdict, returned as UNMEASURED instead of a made-up score.

Most benchmarks hand you a number no matter what. This one refuses. That refusal is the product.

Start here

• New to the tool? The README has the 60-second, no-API-key start.
• Wondering about a result? The FAQ answers the common ones — including "it said UNMEASURED, did it fail?" and "zero measured KEEPs — is the tool broken?" (no on both).
• Want the design? docs/PRD.md is the full specification: evidence model, oracle tiers, admissibility rules, CLI surface.
• Want to see it work? The case studies show the discipline catching its own author before a bad result could ship.

A note on this wiki

This wiki is orientation and FAQ only. The specifications, findings, and case studies live in docs/ in the repository, version-controlled alongside the code — that is the single source of truth. Anything here that conflicts with docs/ is the wiki being stale; trust docs/.