writing-great-skills

Front-load the skill's leading word, list one trigger per distinct branch, and cut identity statements already covered in the body. Every word in a model-invoked description adds permanent context load, so prune it harder than the body.

Split by invocation when you have a distinct leading word that should trigger a model-invoked skill on its own. Split by sequence when visible upcoming steps tempt the agent into premature completion of the current step.

A model-invoked skill keeps its description so the agent and other skills can fire it, paying context load every turn. A user-invoked skill sets disable-model-invocation, making it reachable only by the human typing its name, with zero context load.

This is premature completion, caused by vague completion criteria or visible post-completion steps pulling attention forward. First sharpen the completion criterion to be checkable; only if it stays fuzzy, split the sequence to hide later steps.

Disclosed reference behind a context pointer only loads when the pointer fires, and weak pointer wording makes retrieval unreliable. Inline material every branch needs, and sharpen pointer wording before pulling must-have content back inline.