Agent Skills
A skill is a folder containing SKILL.md (instructions) and optional scripts, templates and reference files. Claude loads a skill only when needed — this gives specialized knowledge without filling the context.
15.1 Progressive disclosure
- 1.Level 1: Metadata: name + description — always in context (very small)
- 2.Level 2: SKILL.md body: loaded when the task matches the skill
- 3.Level 3: Extra files: reference docs, scripts — only when needed; scripts are run, not read
That is why installing many skills does not fill the context
15.2 SKILL.md format
---
name: sql-review
description: Reviews SQL Server queries for performance and security.
Use when the user shares a SQL query, stored procedure or execution plan.
---
# SQL Review
## Steps
1. Check the query for SELECT *, missing WHERE, implicit conversions.
2. Read `reference/indexing.md` for index suggestions.
3. If there is an execution plan, run `scripts/parse_plan.py`.
## Output
Table: Issue | Severity | Fix15.3 Where skills run
| Surface | How |
|---|---|
| claude.ai | Enable/upload skills in Settings; built-in skills (docx, xlsx, pptx, pdf) |
| Claude Code | ~/.claude/skills/<name>/SKILL.md (personal) or .claude/skills/ (project); also via plugins |
| Claude API | In container.skills with the code execution tool; upload through the Skills API |
| Agent SDK / Managed Agents | Attach skills in the agent config |