How scoring works

The full scale has five axes. Only the first two run today, because those two are static analysis — read the file, run nothing, call no model. The other three need the skill to actually run and need real usage data, so on a skill page they read not scored yet, not 0.

325 skills scored. None has reached Verified — that rank requires Effectiveness and Cost to have run.

Safety — 25 points

What the skill does to your machine and your data.

A1 · Destructive commandsMatches rm -rf into a root path or an unchecked variable, DROP TABLE, TRUNCATE, git push --force, git reset --hard, dd if=, mkfs, chmod 777 /, fork bombs, and the curl | sh pattern.6 pts

A2 · Data exfiltrationOnly fires when BOTH appear in the same file: reading process.env, os.environ, ~/.ssh or ~/.aws, and sending to a host outside the allowlist. One half alone does not count.6 pts

A3 · Hardcoded secretsgitleaks-style rules: sk-, ghp_, AKIA, xoxb-, BEGIN PRIVATE KEY, plus entropy on strings of 32 characters or more. URLs and hex digests are stripped before measuring.4 pts

A4 · Prompt-injection guardOnly scored when the skill pulls in web or user content. Full marks if it says to treat that content as data; half if it does not. This is a SIGNAL, not proof.5 pts

A5 · Permission footprintCounts allowed-tools. Unbounded Bash loses the whole criterion. Declaring nothing at all gets half.4 pts

Transparency — 15 points

Whether you can tell what it does before you install it.

B1 · No obfuscationBase64 blocks of 200 characters or more, eval or exec on dynamically built strings, lines longer than 500 characters.5 pts

B2 · Declares dependenciesWhether it states which tools, APIs, models or paid accounts it needs.4 pts

B3 · DocumentedFive items at 0.8 each: purpose · input · output · limits · a runnable example. Broken frontmatter costs points here too.4 pts

B4 · Pinned versionWhether there is a CHANGELOG or a version number, so you know which version you are reading.2 pts

The three that have not run

Effectiveness, Cost and Real-world use need the skill run against a test set, and need install-retention measured over 30 days. Neither exists yet, so we leave them blank. Showing a guessed number here would break the one thing that makes this place different.

The three badge ranks

Unverified — the default. Safety and Transparency only. Deliberately faint.
Verified — Effectiveness and Cost have run.
Gold — 85 or more in total, with at least 30 days of real usage data.

When a newer version of a skill is found, the old badge comes off and is replaced by awaiting rescan. Leaving an old badge on new code is a lie, even for a few minutes.

See the scored skills