Building Great Agent Skills: The Missing Manual

You can watch the video here: https://youtu.be/UNzCG3lw6O0?is=CCpC77FhIZgyZ3S9

Building good AI agent skills sounds tidy until you actually start doing it. Then the mess shows up. A trigger that fires too often. A skill file that keeps growing. Instructions that look important but don’t change anything. The video Building Great Agent Skills: The Missing Manual is useful because it deals with that mess directly.

The core idea is a Skill Checklist Framework, and it starts with a basic question: should a skill be user-invoked or model-invoked? That choice shapes everything. User-invoked skills give you control, but they ask more from the person driving the system. Model-invoked skills feel smoother, though they can add unpredictability and more context pressure. In practice, this is the sort of tradeoff teams run into fast, usually after the first few “why did it do that?” moments.

The next layer is structure. The video recommends splitting skills into two parts, steps and reference. That sounds modest, but it keeps the skill.md file from turning into a storage closet. Big reference sections can move behind context pointers, which helps keep maintenance sane. I’ve seen this pattern in a lot of agent work, once a file becomes the place where everything goes, it stops being useful pretty quickly.

Then comes steering. This is where language matters more than people expect. The video points to leading words, terms that carry a lot of meaning in a small space, and to breaking complex work into smaller skills so the agent has to do more leg work instead of jumping ahead. That’s a subtle but important design move.

Finally, there’s pruning. Keep one source of truth. Remove sediment. Cut no-ops. That part feels unglamorous, but it’s usually where the real quality comes from.

If you’re working on agent systems, this is a practical framework worth keeping nearby. It pairs well with Kun Chen’s agentic engineering workflow, especially if you’re already thinking about how structure and control shape everyday agent performance.

The bigger picture is encouraging: better agent skills probably won’t come from more instructions. They’ll come from cleaner ones.