edram A skill is a progressive-disclosure document set the agent loads on demand. Only the name + description are pre-loaded into context; the SKILL.md body is read when the skill triggers, and references/*.md files are read one at a time only when needed. Authoring a skill means writing for that loading model — not dumping a manual into context.
- SKILL.md is a router, not an encyclopedia. It gives a short overview and points to reference files; depth lives in
references/. - Only add what the model doesn't already know. Every token competes with the conversation. Skip what an LLM is already confident about.
descriptiondrives discovery. Write it in third person, stating what the skill does and when to use it, with concrete trigger terms.- Match this repo's house format, not Anthropic's generic
# headinglayout. Model new skills on the existinggit-commit-conventionskill. - Skills live in
skills/{skill-name}/— theskills/subdirectory, never the repo root (the repo itself is also namedskills).
This skill is pure markdown — no scripts or tooling, matching repo convention.
Core
| Topic | Description | Reference |
|---|---|---|
| Structure | Directory layout, SKILL.md + reference templates, file/skill naming | core-structure |
| Frontmatter | name and description rules — the discovery-critical metadata |
core-frontmatter |
| Workflow | End-to-end process for creating and modifying a skill | core-workflow |
Features
| Topic | Description | Reference |
|---|---|---|
| Progressive Disclosure | Router pattern, one-level references, when to split, TOCs | features-progressive-disclosure |
| Writing Style | Conciseness, degrees of freedom, terminology, examples/templates | features-writing-style |
| Anti-Patterns | Common mistakes + condensed pre-ship review checklist | features-anti-patterns |
| Evaluation | Lightweight, script-free way to sanity-check triggering and value | features-evaluation |