← ClaudeAtlas

writing-skillslisted

Use when adding a new skill or editing an existing one. Keeps the catalog short, the frontmatter correct, and the body free of coercion and padding.
felipemelendez/llm-orchestrator · ★ 3 · Code & Development · score 69
Install: claude install-skill felipemelendez/llm-orchestrator
# Writing skills How to add a skill that fits LLM Orchestrator's bar. ## Before you write Check that an existing skill doesn't already cover the trigger: ``` grep -l 'description:' skills/*/SKILL.md | xargs grep -h 'description:' ``` If something already triggers on this case, edit it. Don't add a near-duplicate. ## File shape ``` skills/<name>/SKILL.md ``` One directory, one file. Optional siblings (helper scripts, prompt fragments) live alongside. ## Frontmatter (two keys) ```yaml --- name: <name> description: Use when <trigger>. <Optional second sentence on scope.> --- ``` Rules the linter (`tests/validate-skills.sh`) enforces: - `name` must equal the directory name. - `description` must start with `Use when`. - The whole file must be ≤ 250 lines (target ≤ 150). ## Body conventions Nothing below the frontmatter is machine-checked, and no fixed section list is required — a previous version of this file mandated five sections in order that no skill in the corpus (including this one) satisfied. The working skeleton, used where each part earns its place: - **One-line purpose** at the top (no header) — near-universal; keep it. - **When to use / when not to** — include when the trigger has a plausible mis-fire; the frontmatter `description` already carries the primary trigger. - **Steps** — numbered, when the skill is a procedure rather than a reference. - **Output shape** — the Concise Agent Protocol block the skill produces, when it produces one. - **Anti-pat