Skill Invocation Trigger

The first design decision for a skill: how it gets invoked, which comes down to a deliberate choice between user-invoked and model-invoked. Any skill can always be invoked manually (the agent reads it on request). A model-invocable skill additionally has a description that always sits in the agent's context; the agent can read that description and decide on its own to pull the full SKILL.md into context — the description acts as a context pointer (Minimal Skill Surface, Context Routing). A user-invocable skill sets disable-model-invocation: true so its description shows only to the user: the model can't see it and therefore can only be told to run it.

The point of the concept is that neither choice is free — they impose different costs, so it's a genuine trade:

Model-invoked → context load + unpredictability. Each such skill adds a description to the agent's context on every request (100 skills = 100 descriptions costing tokens each) plus one more thing for the agent to weigh. And because a context pointer is optional, the model may simply not follow it even when the skill is perfect — which forces you to eval whether skills fire at the right time, a "really nasty" class of problem.
User-invoked → cognitive load on the operator. The more user-invoked skills you have, the more the human must keep in their head and choose to invoke — "more skill required from the pilot."

The source personally prefers user-invoked to stay in full control, keep the agent's context minimal, and remove the fire-at-the-right-time eval problem entirely — accepting more of his own cognitive load in exchange. The comparison he draws: his mattpocock/skills are user-invoked, whereas the rival "superpowers" repo is primarily model-invoked.

Claims

Decide deliberately whether a skill is user-invoked or model-invoked — it's a genuine trade, not a strict win for either. principle — durable: both modes carry real, different costs, so the choice is a design judgment, not a default.
A skill's description is a context pointer: model-invocable means the description sits in the agent's context and the agent may choose to load the SKILL.md; disable-model-invocation hides the description from the agent so only the user can invoke. observation — the invocation mechanism as the source describes it; the disable-model-invocation: true key is a groundable, check-worthy detail.
Model invocation imposes context load: every model-invoked skill adds a per-request description (100 skills = 100 descriptions) plus a decision for the agent. best practice — context: a growing skill library where per-request token budget and agent attention matter; tamp down the count or prefer user-invocation.
Model invocation buys unpredictability — even a perfectly-matched skill may not be invoked — which forces you to eval whether skills fire correctly; user-invocation removes that problem class. best practice — context: an operator who values predictability/control and can absorb the cognitive load of remembering their skills; a team optimizing for hands-off automation might weigh it differently.
User invocation imposes cognitive load: the more user-invoked skills, the more the operator must remember and choose to invoke. observation — the countervailing cost that keeps the trade honest.
Bulk-adopting a third-party skill roster is a bulk purchase of the user-invoked cost, made by default and without deliberation. best practice — context: cloning someone else's packaged agentics (Public Skill Adoption, Role-Typed Agent Roster), where the invocation-mode decision this page insists on being deliberate is instead inherited N times at once from the repo author. The countermeasure is pruning after adoption, not before — you cannot tell which roles fit your work until you've reached for them. Note the trade does not vanish: user-invoked entries stay out of the agent's context, so the bill lands entirely on the operator.
Invocation mode is surfaced to the operator at the command surface, alongside origin — an observed /-palette tooltip reads canary — Post-deploy canary monitoring. (gstack) (user), tagging the originating repo and the user/model invocation mode. observation — a groundable, point-in-time UI detail; useful because it means a roster's invocation modes are auditable after install.
The user-invoked context saving is measurable, and small: /context reports the whole library at 660 tokens for 9 skills (0.1% of the window). observation — the same author's own repo, measured on screen, with his explanation: "not many of my skills actually leech their way into the description, and the descriptions I have are quite short and precise." Useful as the concrete figure behind this page's context-load argument — and note the wrinkle: the installer offered 38 skills while /context counts 9, unexplained on screen. The (user) tag this page records is visible again in the same palette (/grill-me, /grill-with-docs, /grilling).

Minimal Skill Surface — the description-as-context-pointer is the same progressive-disclosure mechanism that keeps SKILL.md small.
Context Routing — "the agent won't load what it isn't pointed to (and may not follow the pointer)" is routing's unpredictability, here at the skill-trigger scale.
Skill Authoring Checklist — this is pillar 1 (Trigger) of the rubric.
Agentic Simplicity — "tamp down the number of model-invoked skills" is the simplicity discipline applied to a skill library's context budget.
Public Skill Adoption — where the invocation-mode decision gets made for you, at clone time.
Role-Typed Agent Roster — N invocation decisions inherited in one git clone.
Skill Pruning — the only remedy once the palette is full.
Context Smart Zone — the budget the description load is actually charged against: not the nominal window, but the usable band you plan inside.
Distillate: Building Great Agent Skills: The Missing Manual
Distillate: 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)
Distillate: The whole flow, end-to-end: the smart zone is the unit of work — the user-invoked claim measured: 660 tokens for 9 skills, and the (user) tag visible in the palette.

Skill Invocation Trigger

Claims

Related

Linked from