← ClaudeAtlas

skill-developmentlisted

How to author a Claude Code Agent Skill — SKILL.md structure, routing-optimized descriptions, progressive disclosure, valid frontmatter fields, bundling scripts/reference files. Use when creating a new skill, writing or fixing a SKILL.md, deciding what belongs in frontmatter, splitting an oversized skill into reference files, or debugging why a skill never triggers.
nxtg-ai/forge-plugin · ★ 5 · AI & Automation · score 73
Install: claude install-skill nxtg-ai/forge-plugin
# Skill Development Procedure for authoring Claude Code Agent Skills. A skill is a directory with a `SKILL.md` file; optional bundled scripts and reference `.md` files sit beside it and load only when needed. ## What a skill is (one paragraph) At startup Claude preloads only the `name` + `description` of every installed skill into its system prompt. When a task matches, Claude reads the full `SKILL.md` body into context. Bundled files (`reference/*.md`, `scripts/*`) are read only when the body points to them. This is **progressive disclosure**: metadata → body → bundled files. Design each level to be the smallest thing that lets Claude decide whether to go deeper. ## Build workflow 1. **Start from a real gap.** Author a skill only for a capability Claude currently fumbles — a workflow it gets wrong, procedural knowledge it lacks, org context it can't infer. Don't pre-write skills for hypothetical tasks. 2. **Create the directory + SKILL.md.** `skills/<kebab-name>/SKILL.md`. Directory name is the on-disk identity — don't rename it after wiring. 3. **Write the frontmatter** (see below). The `description` is the single highest-impact line — write it as a routing rule, not a summary. 4. **Write the body** as *what to do*, not narration. Numbered procedures, concrete examples, a Gotchas section. 5. **Keep it lean.** Target under 500 lines. When it grows past that (or when some content is only relevant in a narrow sub-case), move detail into sibling `reference/*.md` files and