Chapter C53 · Interactive Course
A skill packages reusable know-how
Triggers, entrypoints, staged loading, resources and a small evaluation.
Core Learning Concept
Design a small reusable skill for the stated recurring job. Write a clear description, a focused procedure, and references to any needed resources. Test a matching request, a nearby request that should not use it, and the resulting artifact. Selection and successful execution are separate things to verify.
What You’ll Learn in This Chapter
-
A skill is a reusable task package
Identify instructions and resources for a repeated job.
A skill packages reusable instructions and supporting resources for a class of tasks. It can reduce repeated setup and encode a useful procedure. Installing it does not retrain the model. The host determines how the package is discovered, loaded, and allowed to use tools.
-
The description helps select the skill
Write a trigger that names when the task applies.
A skill description helps a host or model decide when the skill is relevant. Name the task, inputs, and boundaries clearly. “Create a formatted invoice from supplied line items” is more useful than “help with business.” Test nearby requests too, because an overbroad description can trigger unnecessary work.
-
SKILL.md is the entrypoint
Locate required metadata and the procedure.
SKILL.md is a common entrypoint containing metadata and task instructions under the skill format used here. Supporting files can hold references, templates, or scripts. Follow the host’s required structure and validation rules. A correctly named file still needs a useful procedure and resolvable resources.
-
Progressive loading saves context
Explain loading details only when needed.
Progressive loading discloses information in stages: a short summary for discovery, the core procedure when selected, and lengthy reference documents only on demand. This minimizes unnecessary token consumption while preserving critical task details.
-
Resources have different jobs
Distinguish a reference, template and executable script.
References explain, templates provide a starting structure, and scripts execute code. Those resources have different effects and review needs. A bundled script may access files or services with the process’s permissions. Inspect executable behavior and dependencies instead of treating everything inside a skill folder as passive documentation.
-
Invocation depends on the host
Distinguish automatic selection from an explicit request.
Invocation is how a skill is selected and loaded. A host may support automatic matching, explicit naming, shortcuts, or a combination. Packaging the procedure does not guarantee every invocation mechanism exists. Check the current host’s documented behavior and verify selection using an appropriate test request.
-
A skill needs a small evaluation
Test triggering and the delivered result separately.
A small skill evaluation tests whether it is selected appropriately and whether it produces the intended result. Include a relevant request, a nearby out-of-scope request, and a difficult valid case. Inspect the final artifact and side effects, not just the agent’s statement that it followed the skill.
-
Draft a tiny skill
Prepare a fictional meeting-summary skill with a precise trigger, short procedure, one template, an irrelevant-request case and an output check.