Skip to skills
AI Craft Pad
Home/Skills
Transparent methodology

How the AI Craft Pad Skill Score works

A discovery aid built from observable signals. Every signal has a scope, and unknown means unknown.

Score version 1.0

The total is the rounded weighted sum of eight 0–100 evidence signals: documentation 20%, repository maintenance 15%, license clarity 10%, repository health 10%, compatibility evidence 10%, repository popularity 5%, originality 10%, and static instruction-text review 20%.

Documentation uses bounded counts of words, headings and examples. Maintenance uses the repository push date: 100 within 30 days, 75 within 90, 40 within a year and 10 beyond a year. Archived or unavailable sources receive zero maintenance points. Stars use a logarithmic repository scale. License clarity means a declaration was found, not that all bundled resources share that license.

Safety stays provisional

The static scan looks only at SKILL.md text for remote shell pipelines, destructive commands, credential references, network access, package installation and selected prompt injection phrases. It does not execute code, inspect dependencies, prove intent, or certify a skill. Pattern matches can describe legitimate examples. A clean text scan earns at most 60/100 for this signal; high-risk patterns earn less. Runtime safety remains Unknown.

Compatibility and originality

Agent tags reflect explicit ecosystem ownership or repository README statements. They are repository-level evidence, not successful tests of every skill on every agent. Identical normalized instruction bodies are grouped by hash. A first-party/non-fork source is preferred within a matching group, but this is not proof of authorship. Fork status and duplicate content are shown separately.

Indexing and freshness

Most metadata records are useful inside this search but remain noindex. A detail page needs a substantive original editorial review bound to its instruction hash and exact source path, valid format, source commit, declared license, current metadata and no duplicate, deprecated, archived, unavailable or high-risk instruction flags. Changed instructions invalidate the old review until it is checked again. A high numeric score cannot unlock indexing. Query and filter URLs stay noindex and point to the clean collection canonical.

New means first seen in this catalog. Trending is measured repository star growth between comparable snapshots at least 20 hours apart; first-sync entries have no trend value. Repository push time is not an individual skill update timestamp.