Teach a reusable method
Use when authoring or evaluating a skill. Use current tool descriptions for storing, updating, binding and loading it; do not copy their API contracts into the skill.
Start with repeatable decisions a future agent cannot infer from a tool schema. A skill is useful know-how, not a transcript, a fact store, an inventory of tools or a grant of authority. Inspect existing methods before adding another.
Write a precise description that selects the skill for the right requests and excludes nearby routine operations. Keep its body short: purpose, decisions, failure handling, quality checks and the evidence needed to finish. Put examples, templates and lengthy references in separately named resources loaded only when necessary.
Extract general lessons from successful work while removing customer-specific identifiers, correspondence, credentials and incidental history. Examples must be appropriate for the skill's audience. A shell example does not establish installed dependencies.
Evaluate an ordinary request, a near-match that should not load the skill, missing inputs and hostile source instructions. Judge observable effects and output quality. Revise unnecessary or conflicting instructions. Preserve organization-authored changes when updating a built-in; installing or saving a skill does not itself bind it to every agent.
