evaluating-skillslisted

# Evaluating Skills Skill development has two phases: **drafting** (`slow-powers:writing-skills`) and **evaluation** (this skill). This skill owns the *craft* of evaluation — deciding whether a change needs measuring, designing test cases, devising pressure-testing scenarios, writing assertions, and reading results. The *mechanics* of actually running an eval — building the workspace, staging skills, dispatching subagents, grading, aggregating — are owned by a dedicated tool, **[eval-magic](https://github.com/slowdini/eval-magic)**, which ships as a dependency-less prebuilt binary you invoke as `eval-magic`. See [Running the eval](#running-the-eval) for the hand-off. ## Overview An eval is a structured measurement of whether a skill actually shifts behavior. Each test case is a realistic prompt; each run dispatches a fresh general-purpose subagent twice — once with the skill loaded, once without (or once with the prior version, once with the revised version) — and grades the outputs against assertions. Pass-rate deltas tell you whether the skill is worth shipping or the change is worth landing. Evals are harness-agnostic: run records use a portable JSON schema so an eval authored on one harness (Claude Code, Codex, OpenCode) can be executed and graded on any other. The runner documents the schema and the per-harness specifics; this skill stays at the level of *what makes a good eval*. ### Two comparison modes - **Mode A — new skill.** Compares `with_skill` vs `without_s