Skip to content

Importing a skill

Most people writing a Custom Skill have already written most of it somewhere else: a brand guide, a shot spec, a prompt document, a skill for another tool. Importing turns that into a Skill rather than making you retype it into a form.

Give it the source and it fills what it can. Seven fields:

Field What it takes from your source
Name A short title.
Routing Which requests should route here, with trigger phrases in your own words. An Anthropic SKILL.md’s frontmatter description lands here.
Composition rules What holds in the frame every time (subject placement, light, materials, palette, framing), stated as rules a model can execute.
Technical constraints Resolution, text, format, color and material realities; deliverable specs.
Aspect ratios Only what the source states or clearly implies. Empty otherwise, rather than guessed.
Clarifying questions What to ask before making the work.
Failure modes Each as a failure, its mechanism, and a correction that says what to specify.

Then you confirm the routing before it saves. The field that decides whether the Skill ever fires is never left as the import guessed it.

The honesty report is the part worth reading. Three fields are deliberately not filled by the import: intent, the model fit, and the model translation notes. When part of your source belongs to one of those, the import says which field it belongs to rather than reporting that no field holds it.

That distinction sounds small and is not. An import that says “nothing here holds this” teaches you to throw the material away. One that says “this belongs to intent, which you set afterwards” teaches you where to put it. The first version of this import had the earlier behavior, and it was reporting honest loss about content the record could perfectly well carry: the type was right and the instructions were wrong.

intent is derived if you don’t set it, from the first sentence of your composition rules. Worth checking, since it is the one-sentence statement of what the Skill is for and the first sentence of your rules may not be it.

Aspect ratios stay empty rather than becoming a guess. Only what your source states or clearly implies. An invented ratio is worse than none: it silently constrains every generation the Skill touches.