| name | skill-generator |
| description | Use when asked to create, generate, author, or bootstrap a NEW skill from a natural-language description. Turns "make me a skill that does X" into a stored SKILL.md by drafting valid frontmatter + body and writing it with the existing write_skill tool — dogfooding the skill system. |
Skill Generator
Generate a new, reusable skill from a description and store it. The skill system
already owns storage (write_skill → ~/.vanta/skills/<slug>/SKILL.md); this
skill is the authoring discipline that feeds it. Do NOT hand-roll a file write.
When to use
The user describes a capability they want captured as a skill: "create a skill
that summarizes PRs", "make a skill for our deploy checklist", "bootstrap a skill
from this description".
Steps
- Read the description. Extract the skill's purpose and WHEN it should fire.
- Draft three fields:
name — a short kebab-case slug (lowercase, hyphens, no spaces). This
becomes the directory; keep it slug-safe (no /, no ..).description — ONE sentence stating when to use the skill. This is what
recall matches against, so lead with the trigger ("Use when …").body — a Markdown how-to: a # Title, then concise numbered steps or
rules. Write the procedure, not prose about it.
- Write it with the existing tool. Call
write_skill with
{ name, description, body } (and optional tags). That persists and commits
the SKILL.md and wires it into the registry — it appears in /skills and is
recallable next session. Never write the file by hand.
- Confirm. Report the stored slug and path from the tool result.
Rules
- One concern per skill. A description spanning several jobs → several skills.
- The
description is the index entry — make it specific and trigger-first.
- The
body must let an agent reproduce the behavior WITHOUT re-asking. No TODOs
or placeholders.
- Reuse existing skills before generating a near-duplicate (
recall first).
Reference
write_skill tool — the storage path (do not reimplement it).skills/skill-gen.ts — pure buildSkillFromDescription /
parseSkillModelOutput (prompt builder + JSON→{name, description, body}
validator, kebab-validated so the path can't escape) if generating
programmatically rather than authoring inline.skills/store.ts writeSkill — preserves created, bumps updated.
Verification
- After
write_skill, the new skill shows in /skills and recall finds it by
its description.
- The stored frontmatter has
name, description, created, updated, tags.
