skills_wiki
#ranking.md

How the S/A/B/C rank is computed

Every skill is scored by code, not by an AI's impression of it. First, a set of structural gates check whether the skill is actually well-formed. If it passes all of them, five objective signals combine into a 0–100 score that determines the letter rank.

1. Structural gates (all must pass)

Failing any gate caps the skill at rank C. This is the check that catches broken installs — skills that look fine at a glance but resolve to nothing at runtime.

Valid frontmatter

The skill has a real name and description — not a stub or a heading mistaken for a skill.

Not a boilerplate heading

The skill's name isn't a repo doc-noise heading like "license" or "getting-started" — a marker of a broken ingest, not a real skill.

Substantial content

The skill body has enough real instruction, not just a title.

No unguarded destructive commands

No unsafe patterns like an unguarded rm -rf or DROP TABLE.

Referenced files exist

If the skill points to scripts or reference files, those files actually exist in the repo.

2. Weighted signals (gates already passed)

These combine into a 0–100 composite score.

25%

Documentation

Clear description, a stated trigger condition, and a usage example.

25%

Structure

Organized into headings and steps rather than one wall of prose.

20%

Resources

Referenced scripts and reference files are real and resolvable.

20%

Community

GitHub stars on the source repository — a real but bounded signal, not the whole score.

10%

Recency

The source repository has been maintained recently.

3. Rank thresholds

S85+

Passes every gate and scores well across all five signals.

A65–84

Passes every gate; solidly documented and structured.

B40–64

Passes every gate but is thin on documentation, structure, or community signal.

Cunder 40, or any gate failed

Structurally broken, unsafe, or too thin to be a real skill.

What this rank isn't

This rank measures structural soundness, documentation completeness, and community trust signals — not whether the skill's actual advice is good. A skill can be S-ranked (well-formed, well-documented, verified references, a starred repo) and still give mediocre guidance for its specific domain. Treat the rank as an honest floor on "is this a real, complete, safe skill" — not a guarantee of the quality of its advice.

Skills with many sub-skills

Some packs bundle many sub-skills from one source repository instead of shipping a single skill. For those, the same gates and signals are applied to a sample of the sub-skills (up to 40, evenly spread across the collection) and combined with the median — one bad entry in a large collection doesn't sink the whole pack's rank, but a pack where most sampled sub-skills fail the gates still lands at C. The skill's detail page notes how many sub-skills were scored.

← Back to marketplace
// isolated_environments: Railway private containers// performance: FastMCP 3.0, sub-second latency// evolution_loop: Gemini-powered auto-patching