菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Route a task to the least intrusive executive-function support based on the user's stated current capacity, memory load, activation, and motivation. Use when work feels stuck, overwhelming, hard to hold in mind, or when the user asks for neurodivergent-friendly support without wanting a diagnostic or coaching ritual.
Reduce work to the smallest safe and useful outcome when capacity is low or burnout is present. Use when the user explicitly says they are depleted, overwhelmed, unable to start, or asks for the minimum viable version of a task.
Create a truthful activation bridge for a task that has low urgency, novelty, interest, challenge, or immediate feedback. Use when the user says they are stuck or avoiding a known task despite understanding it.
Reduce language, sensory, and uncertainty load while preserving complete information. Use when the user requests direct or scannable communication, has low capacity, is navigating a new flow, or needs explicit expectations and examples.
Turn a vague, oversized, or initiation-resistant outcome into an executable sequence with a visible first action and definition of done. Use when dependencies are hidden, the task spans multiple systems, or the user cannot identify where to begin.
Externalize time, checkpoints, stopping rules, and context transitions. Use when estimates are uncertain, a task may expand indefinitely, the user is switching topics, or an interruption risks losing the current state.
| 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. |
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.
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".
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_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.description is the index entry — make it specific and trigger-first.body must let an agent reproduce the behavior WITHOUT re-asking. No TODOs
or placeholders.recall first).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.write_skill, the new skill shows in /skills and recall finds it by
its description.name, description, created, updated, tags.