Quote Card

Claude Code · Skill Design

Make a skill model-invocable and its description stops being documentation — it becomes the activation trigger. Word it loosely and it fights built-in Plan Mode for the same intent.

agentics · plan-agent plugin

One flag flip taught me the subtlest rule in Claude Code skill authoring.

A skill in our plan-agent plugin was manual-only — it ran only when you typed the slash command. We wanted Claude to reach for it automatically. The fix looks like a one-line deletion: drop `disable-model-invocation: true`.

It isn't that simple. The moment a skill becomes model-invocable, its `description` field stops being documentation and becomes the activation trigger. Word it loosely and the skill fights Claude Code's built-in Plan Mode for the same "plan this" intent.

Three things made it safe:

1. Scope the trigger to concrete artifact nouns — "create an HTML plan", "implementation-plan document" — not generic verbs like "plan how to X". Specific nouns win; broad intent collides.

2. Make argument-sourcing dual-mode. The command path still parses $ARGUMENTS; the model path reads the objective from the conversation and only asks when nothing can be inferred.

3. Ship it as a MINOR bump — a new activation path on top of an unchanged command is purely additive.

The lesson: for a model-invocable skill, precision in the description is everything. It's not where you explain the skill — it's where you decide when it fires.

Follow for more Claude Code plugin patterns like this.

#ClaudeCode #AITools #DeveloperTools #Plugins

Twitter/X

Making a Claude Code skill model-invocable is nearly a one-line change: delete `disable-model-invocation`.

But that line turns your `description` from docs into the activation trigger. Scope it to artifact nouns, not generic verbs, or it fights built-in Plan Mode.

#ClaudeCode

Bluesky

Subtle Claude Code lesson: making a skill model-invocable is a one-line flag deletion — but it flips the `description` field from docs into the activation trigger.

Scope it to concrete artifact nouns ("create an HTML plan"), not generic "plan how to X", or it collides with built-in Plan Mode.

Substack

A subtlety worth sitting with: in Claude Code, the moment you make a skill model-invocable, its `description` field changes jobs. It stops being where you document the skill and becomes where you decide when it fires.

A vague description isn't just unhelpful — it's an activation bug. Scope it to concrete artifact nouns and it cooperates with built-in Plan Mode; leave it generic and the two compete for the same intent.

Follow for more Claude Code patterns.