Video: "How AI agents & Claude skills work (Clearly Explained)" by Ras Mic / Clearly Explained on YouTube.

What Skills actually are

A Skill is a Markdown file — stored in a .claude/skills/ folder in your project — that describes how to handle a specific class of task. Think of it as a standing instruction: "when you need to write a release note, use this format and run these checks." The agent loads the full file only when it decides the Skill is relevant, then sets it aside once the task is done.

That on-demand loading is the important bit. Without Skills, teams tend to dump everything into a single CLAUDE.md that grows unwieldy over time. With Skills, each playbook stays separate and focused, appearing in context only when called upon.

Why the token saving is worth understanding

Token costs are easy to ignore until they're not. The video cites a comparison worth noting: a Skill costs roughly 53 tokens per turn, while an equivalent block sitting permanently in CLAUDE.md costs over 900. That gap compounds quickly across a team running dozens of tasks a day.

To be fair, the saving only applies if you actually split your instructions into Skills rather than piling more into a monolithic config file. Keep the big CLAUDE.md and add a Skills folder on top of it, and you don't save anything.

What's overhyped

The framing that Skills make Claude "fully autonomous" is doing a lot of heavy lifting. Skills make the agent more consistent on tasks it already does well. They do not compensate for vague instructions, poorly scoped tasks, or a codebase the agent hasn't been oriented to properly. You still need to do the groundwork first.

What's genuinely useful in practice

The advice on how to create Skills is the most practical takeaway from the video. Rather than writing a Skill from scratch, walk through the task with the agent until you get a result you're happy with — then ask it to write the Skill based on what it just did. That way the Skill reflects your actual workflow rather than what you think your workflow is. It's a small shift in approach but produces noticeably better results.

Candidates for Skills in a typical business: brand-voice copy checks, PR description templates, invoice rewriting, SEO audit checklists, deployment runbooks. Anything you find yourself explaining to the agent more than twice is worth turning into a Skill.

Where this connects to NordSys

Setting up Skills well is exactly the kind of thing that sounds straightforward and then takes a day when you try it yourself. Getting the scoping right, deciding what belongs in a Skill versus a hook versus your main config, and keeping the whole setup tidy as your workflows evolve — that's what we handle on NordSys AI Agents care plans. We build the Skills for your team's actual tasks, not hypothetical ones, and we update them when your processes change.

See care plans for Claude Code →