When not to auto-invoke a skill: disable-model-invocation and paths

Most Agent Skills are meant to wake up when the task matches their description. That is useful until a skill that deploys, commits, or spends money also wakes up because a sentence looked similar. On 2 October 2026, Cursor's Agent Skills docs make the escape hatch explicit: set disable-model-invocation: true so the skill only loads when you type /skill-name, and use paths when guidance should only appear for matching files.

Demand for this corner of the skill ecosystem is real but focused. In the past week, X recent-post counts for agent/cursor/claude skills language totalled about 3.7k posts (non-retweet query checked 2026-10-02). A narrower query tying skills to Cursor/Claude and invoke/auto language was about 770. That is enough signal for a control guide, not a hype roundup.

What auto-invocation actually does

Cursor discovers skills from project and user skill directories (.cursor/skills/, .agents/skills/, and the Claude/Codex compatibility paths). The agent sees available skills and decides relevance from context — mainly the skill description. Manual /skill-name still works for any skill. Auto-invocation is the default.

That default is right for research, drafting, and checklists. It is wrong for side-effect workflows where a false positive is expensive.

Use disable-model-invocation: true for side effects

Cursor documents the field in SKILL.md frontmatter:

---
name: deploy-staging
description: Deploy the app to staging. Invoke only when the user explicitly asks to deploy.
disable-model-invocation: true
---

When true, the agent will not auto-apply the skill from context. You must invoke it with /deploy-staging (or attach it as a Custom Mode for a whole session). Cursor's own /migrate-to-skills converts slash commands into skills with this flag set, so migrated commands keep their explicit-only behaviour.

Good candidates for slash-only:

1. Deploy, publish, or ship — anything that changes a live environment.
2. Commit, open a PR, or merge — git and review side effects.
3. Send a message or charge money — external actions you want to approve every time.
4. Destructive cleanups — delete, reset, force-push helpers.

Keep auto-on for skills that only advise, draft, or inspect without writing irreversible state.

Use paths when the skill is file-shaped

paths (glob list or comma-separated string) limits when the skill is surfaced. Cursor only offers it while the agent works with matching files. Nested project skills under e.g. apps/web/.cursor/skills/ are automatically scoped to that directory — similar to setting paths by hand.

Example:

---
name: react-component-patterns
description: Conventions for React components in this codebase.
paths:
  - "**/*.tsx"
  - "packages/ui/**/*.ts"
---

Use paths when the skill is true only for certain languages or packages. Do not use it as a substitute for disable-model-invocation on deploy skills: file scope still allows auto-invocation inside that scope.

A short decision tree

Write the trigger into description either way. Auto-on skills need a sharp description so they do not steal unrelated turns; slash-only skills still need a clear description so humans can find them in /.

Skills already on Skill Rank that help

These are already listed on Skill Rank — no catalog add required for today's post:

• using-agent-skills (addyosmani/agent-skills) — how to live with the skill model day to day.
• write-a-skill (mattpocock/skills) — authoring habits that keep triggers honest.
• skill-creator (anthropics/skills) — Anthropic's SKILL.md authoring workflow.
• writing-skills (obra/superpowers) — another pass on scope and clarity.

If you are tightening an existing skill, start by editing frontmatter before rewriting the whole body: one boolean and a few globs often fix noisy auto-invocation.

Limits of this note

• Field names and behaviour follow Cursor docs checked 2026-10-02; other agents may use different activation rules (the Agent Skills spec discussion around standardising disable-model-invocation is still open).
• X counts are recent-search volume, not Google search demand or Skill Rank rankings.
• Stars on a catalog row do not prove a skill is safe to auto-invoke — read the skill, then decide the frontmatter.