08 · Skills Overview
A skill is a self-contained, portable unit of capability: a SKILL.md manifest with YAML frontmatter, optionally accompanied by scripts/, references/, and assets/. The generated distribution index records the inventory selected by skill-system.json. The UI/UX routing catalog contributes 41 entries, with a 40-entry portable subset shared with mini; its native Impeccable integration is full-only. Use the generated catalogs for current inventories and the UI/UX guide for selective loading. This page explains the model: how skills are structured, how they are discovered, the standard they conform to, and where to find each category.
The skills model
Every skill follows the same shape, which is what makes them deliverable to the manifest-selected AI tools.
skills/<category>/<skill-name>/
├── SKILL.md ← manifest: YAML frontmatter + instructions (under 500 lines)
├── scripts/ ← executable code (optional)
├── references/ ← detailed docs, loaded on demand (optional)
├── assets/ ← templates, schemas, examples (optional)
└── templates/ ← Tera (.tera) code-generation templates (forge-rs skills)
The frontmatter is the contract. A name (lowercase, hyphens, matching the directory) and a description (1–1024 characters, searchable) are required. For new skills, strict validation also requires license, version, and a non-empty metadata.tags array.
---
name: axum-patterns
description: Production Axum web API patterns — routing, middleware, state, extractors
license: MIT
metadata:
author: Prometheus AGS
version: '1.0.0'
category: rust
tags: [rust, axum, web, api, middleware]
---
Two design principles govern every skill. Progressive disclosure: the main SKILL.md stays under 500 lines, and detail lives in references/ that is loaded only when needed — this keeps context lean. Third-person, imperative voice: "Run the command," not "you should run." These are not style preferences; they are what keeps a large skill library usable inside a finite context window.
How skills are discovered and triggered
Skills surface differently depending on the AI tool, but the mechanism is consistent: a skill's description and metadata.tags are what the tool matches against the user's intent.
In Claude Code, the flat installer turns each skill into a slash command (/kbd-init, /evolve, /refine-ui, and so on) and the description drives auto-triggering. The native agent generator's own skills engine uses TF-IDF selection over the configured skill directories to pick the most relevant skill for a task — the same lightweight, explainable matching the Karpathy KB uses, applied to skill selection. Discovery per tool is covered on the Platform Support page.
The AgentSkills.io standard
Native skills are authored against the AgentSkills.io specification. That conformance is what makes a skill portable to any platform that reads the standard — not just Claude Code, but Kimi, MiniMax, OpenCode, Codex, Cursor, Windsurf, Gemini CLI, Roo Code, and Amp. The repository's validator (scripts/validate-skills.js) enforces the spec: frontmatter schema, the name pattern ^[a-z0-9]+(-[a-z0-9]+)*$, name/directory consistency, forward-slash paths only, script executability, and — in strict mode — the presence of license, version, and metadata.tags. The validation gates are documented on the Contributing page.
The full category index
Use the generated Skills Index in SKILLS.md and the site catalog for the complete categories and counts. They are regenerated from actual source at the completed-phase boundary. Hand-maintained counts are omitted here to avoid competing inventories.
The guide covers process skills, language and domain guidance, learning, artifact refinement and UI/UX routing. Former sync skills are outside the current pack's service ownership.
Native, imported, and forge-rs skills
There are three kinds of skill in the repository, and the distinction matters for how you maintain them.
Native skills live directly under skills/<category>/ and are maintained in this repository. Most are uniform v1.0.0, MIT-licensed, with metadata.tags. (One exception worth knowing: kreuzberg is Elastic-2.0 licensed, not MIT.)
Imported skills live under skills/imported/ as git submodules, because they have independent development lifecycles. The two current imports are artifact-refiner and sycophancy-correction. You never edit an imported skill in place — you update its submodule pointer. (See Contributing.)
forge-rs skills carry .tera templates under templates/. forge-rs scans skills/<language>/<skill-name>/templates/*.tera, and each skill's skill.toml declares which templates it ships. The four template variables (task_description, task_id, constitution_summary, karpathy_focus) are filled at enrichment time. This is how a skill contributes not just guidance but actual code scaffolding — covered on the Rust Toolchain page.
The pack also generates new skills on demand. pmpo-skill-creator produces production-ready skills via the PMPO loop in four modes (create, clone, extend, update), while Forge enriches tasks with registry skills and constitution context. Forge has no template-management CLI subcommand. Both are documented in the pages ahead.
Previous: ← 07 · Sycophancy Correction · Next: 09 · Process & Orchestration Skills →