disable-model-invocation: true in a skill's front matter stops the agent from loading that skill on its own: it runs only when you type /skill-name. In Claude Code and Cursor the skill also stays out of the agent's context until you call it. Of 1,972 public skills we read on November 17, 2026, 36 set it, from five repositories.
Use it for a skill whose timing you want to control: a release, a deploy, anything that posts to GitHub or an issue tracker. Claude Code's documentation puts it plainly: "You don't want Claude deciding to deploy because your code looks ready."
Which public skills set disable-model-invocation?
We read the front matter of every skill in this directory's November 16 crawl, 1,972 SKILL.md files from 47 GitHub repositories, cloned and parsed by script on November 17, 2026. Each skill's agents/openai.yaml, where one exists, was read too.
| Setting | Skills | Repositories |
|---|---|---|
disable-model-invocation: true |
36 | 5 |
disable-model-invocation: false (the default, written out) |
0 | 0 |
user-invocable: false (only the agent can start it) |
2 | 2 |
user-invocable: true (the default, written out) |
78 | 1 |
Both disable-model-invocation: true and user-invocable: false |
0 | 0 |
allow_implicit_invocation: false in agents/openai.yaml (Codex) |
24 | 2 |
Source: Use This Skill directory (data/skills.json, crawl of November 16, 2026) and the weekly scan of the same date (data/scan.json). All 1,972 SKILL.md files could be opened and all parsed. The groups below are our reading of each skill's description.
The 36 come from mattpocock/skills (23), Donchitos/Claude-Code-Game-Studios (8), trailofbits/skills (3), tester-army/e2e (1) and tobi/qmd (1). In the weekly scan, 35 are clear and one, firebase-apk-scanner, is marked review because a script prints a sudo apt install line. None is flagged.
Reading the descriptions, the 36 fall into four groups:
| What the skill does | Skills | Examples |
|---|---|---|
| Publishes or changes something shared: GitHub, an issue tracker, a release | 8 | github-triage (can merge pull requests), release, hotfix, to-spec, to-tickets, triage |
| Starts a long, multi-step job you choose when to begin | 14 | implement, dev-story, chief-of-staff, claude-handoff, launch-checklist |
| A conversation mode you start on purpose | 11 | grill-me, teach, retro, wait-what, writing-shape |
| An old name that forwards to a newer command | 3 | changelog, help, settings |
So authors use the field for two reasons: side effects, which is what the documentation describes, and skills that make no sense unless you asked for them. A "relentless interview" about your plan is not something you want to begin because a sentence happened to match.
How do Claude Code, Cursor, Codex and Gemini CLI treat it?
The field is not in the Agent Skills specification, which defines six front matter fields (name, description, license, compatibility, metadata and allowed-tools). Each agent decides for itself whether to read it. Checked on each official page on November 17, 2026:
| Agent | Reads disable-model-invocation? |
What happens | Manual-only switch if not |
|---|---|---|---|
| Claude Code | Yes | Claude cannot invoke the skill; the description is not in context; the skill is not preloaded into subagents and does not run when a session-scoped scheduled task fires with it as the prompt | Also skillOverrides: "user-invocable-only", or a Skill(name) deny rule |
| Cursor | Yes | "Only included when explicitly invoked via /skill-name"; the agent will not apply it from context |
|
| VS Code (GitHub Copilot) | Yes | Agent will not load it automatically; it runs from the / menu |
|
| Codex | Not mentioned | Codex chooses skills from the description | allow_implicit_invocation: false under policy in agents/openai.yaml; $skill still works |
| Gemini CLI | Not mentioned | The agent asks your permission each time a skill is triggered | /skills disable <name> turns the skill off |
This is why the pairing matters. All 24 skills in our count that set allow_implicit_invocation: false for Codex also set disable-model-invocation: true, so they are manual-only in Claude Code, Cursor, VS Code and Codex. The other 12 manual-only skills (all 8 from Claude-Code-Game-Studios, 2 from trailofbits, and the e2e and qmd skills) have no Codex setting, so Codex can still pick them on its own. The Donchitos repository targets Claude Code by name, so for its own users that may not matter. For a skill meant to travel, it does. More on what each agent reads is in do agent skills work in Codex, Cursor and Gemini CLI?
Do skills with side effects set it?
Not usually. We searched each SKILL.md for commands that write somewhere other than your files: git commit, git push, git tag, gh pr create or merge, gh issue create, close, edit or comment, gh release create, npm publish and common deploy commands.
| Command found in SKILL.md | Skills | Of them manual-only |
|---|---|---|
| Any of the commands above | 27 | 5 |
git commit |
10 | 1 |
git push |
9 | 1 |
Deploy (vercel deploy, kubectl apply and others) |
6 | 0 |
gh issue create, close, edit or comment |
4 | 2 |
git tag |
4 | 1 |
gh pr create or merge |
3 | 1 |
A match does not mean the skill runs the command: many of the 22 others show it as an example, a CI template or a warning. Some do run it, and guard it in the text instead. deploy-to-vercel tells the agent to "always deploy as preview" unless you ask for production and to ask before pushing; superpowers' finishing-a-development-branch asks you to confirm before merging. That works when the model follows the instruction. The front matter field is enforced by the agent itself, before the skill is loaded at all.
What does a manual-only skill cost when you call it?
Nothing until you call it, in Claude Code and Cursor. Claude Code's documentation says every skill in the listing adds to your context on every turn, and a manual-only skill is left out of that listing. The 36 descriptions together come to 4,984 characters, median 104.
When you type the command, the whole SKILL.md loads. We have side-by-side proofs of two of the 36, run on claude-opus-5-5 with three runs a side on October 24, 2026, with the skill loaded as an agent would read it once invoked. triage added 6,758 input tokens per run (472 without, 7,230 with); median output was 17,079 tokens without and 18,159 with, and median time 146 and 155 seconds. improve-codebase-architecture added 4,744 input tokens (850 without, 5,594 with) and shortened median output from 32,503 to 28,181 tokens. Both are on the proven list. For why one run is not enough to judge a skill, see how to test whether a skill helps.
Is disable-model-invocation a safety setting?
No. It decides when a skill starts, not what it may do. Once you type /skill-name, the instructions, any scripts and any allowed-tools grant apply in full; in Claude Code that grant lasts for the turn that invoked the skill. A skill can also be manual-only and still be the one you should not install. Check it first: paste its GitHub link into the live skill check, or read what a malicious skill can do.
When should I add it to my own skill?
- Add it when running the skill at the wrong moment would cost you something: a deploy, a release, a merge, a message, a ticket on a shared tracker.
- Add it when the skill is a mode you enter on purpose, such as an interview, a retrospective or a teaching session.
- Pair it with
allow_implicit_invocation: falseinagents/openai.yamlif the skill should be manual-only in Codex too. - Leave it off for knowledge the agent should pick up when relevant, and write a clear description instead. If your skill does not trigger when you expect, the field is one of the causes listed in why a skill does not trigger; the others are in how to write a skill description.
We did not run any of the 36 skills for this page. The counts come from reading their files; the token figures come from proofs already published.