← ClaudeAtlas

writing-skillslisted

Writes or tightens a Claude Code skill: short description, tight steps, tested triggers. Use when the user asks to create, write, improve or shrink a SKILL.md or a project skill, or says a skill never fires. Not for using a skill.
t8jwscc7qy-sketch/foma-skills · ★ 2 · AI & Automation · score 78
Install: claude install-skill t8jwscc7qy-sketch/foma-skills
Writes a skill that is short, fires when it should, and changes behaviour the model would not show unaided. Draft, test, then keep. ## Steps 1. *Always.* Ask whether a skill is the right layer: a one-sentence fact goes to a CLAUDE.md line, a mechanical rule to a hook, a path-bound rule to `.claude/rules/`. A skill is for a repeated multi-step procedure or a method. Say which layer and why. 2. Read what is there: existing skills' names and descriptions (collisions), the CLAUDE.md contract. For a rewrite, read the whole source skill first and keep every line that changes behaviour. 3. Draft `.claude/skills/<name>/SKILL.md` (kebab-case name = folder). Frontmatter: `name`, `description`, optional `license`; add `disable-model-invocation: true` only for skills with side effects that must run on request. No `model`, no XML in the description. 4. Description, at most 250 characters: what it does, then "Use when" with 2-4 situations in the user's own words, then "Not for" the nearest neighbour. Scope by situation, not by naming other skills. 5. Body in this order, empty parts dropped: one-line purpose; Steps (numbered, imperative, each checkable; the first steps marked *always* are the short path for small tasks); Rules (at most 5, "never X: reason"); Branches ("If <condition>, read `references/<x>.md`", one level deep); Report (at most 5 lines, last line "Not checked: …"). Read paths and commands from the project's CLAUDE.md instead of hard-coding them. 6. Cut: rationale, history,