Beat a skill

Read theirs, build ours from scratch, measure both the same way. Better means better on the same tasks with the numbers to show it, not longer.

The engine runs as npx -y @thenavidm/skill-builder@latest. Where skill-builder is installed (npm install -g @thenavidm/skill-builder), run skill-builder instead: it's the same engine, without the download.

1. Get it and read its licence

From GitHub, a skills directory, their disk, or pasted text. Read the licence before anything else and tell them what it allows: reading it to learn is always fine; reusing its code needs a licence that permits it and a credit; copying its text is never what we do, whatever the licence.

2. Take it apart

The skill-architect agent can do this teardown in a fresh context.

• What it does, step by step.
• What it does well: list every strength. Ours must match each one or beat it, or the comparison isn't honest.
• Where it stops: tools it doesn't work in, steps it skips, claims it doesn't test.

Run the engine on it, since numbers are fairer than impressions:

npx -y @thenavidm/skill-builder@latest check <theirs>
npx -y @thenavidm/skill-builder@latest tokens <theirs>

3. Build ours

With skill-create, from the job, not from their text. Close the window with their skill before writing, so the words are ours.

4. The same tasks, both skills

Write the tasks once, covering what both promise, with real inputs. Run each skill on them in the same tools:

npx -y @thenavidm/skill-builder@latest eval <ours> --runs 3
npx -y @thenavidm/skill-builder@latest eval <theirs> --runs 3

Add trigger on the same prompts if triggering matters, and bench if they use several tools. Have the skill-grader agent grade every run. A blind comparison helps when assertions can't capture quality: give the skill-comparator agent the two outputs for each task, unlabelled. Where an AI tool can't run headless, the skill-executor agent runs each task in a fresh context instead.

5. Prove it's ours

npx -y @thenavidm/skill-builder@latest overlap <ours> <theirs>

It lists every run of eight or more words the two share. Rewrite any in our own words and run it again until there are none.

6. Report

A table: each strength of theirs and where ours has it; where ours goes further; the numbers for both on the same tasks; the overlap result. Say plainly where theirs still wins.