If your feed looked like a firehose of "build agent skills from scratch" clips this week, there was a reason: DeepLearning.AI and Anthropic released a free short course, Agent Skills with Anthropic, taught by Anthropic's Elie Schoppik, and Andrew Ng's announcement sent it straight to the top of the founder timeline. Here is the version a team of one can act on before the tab closes.

The one-sentence answer: an agent skill is a folder of instructions an agent loads on demand — and the single line that decides whether it ever loads is the one part almost everyone writes carelessly.

What a skill actually is#

Strip away the branding and a skill is a directory with a SKILL.md file. The top of that file is YAML frontmatter with two fields that matter — a name and a one-line description — and below it is plain markdown: the procedure, the house style, the checklist, the code the agent should follow.

my-skill/
  SKILL.md          # name + description + instructions
  reference.md      # optional: loaded only if the skill needs it
  scripts/build.py  # optional: code the skill can run

The trick the course keeps returning to is progressive disclosure. The agent does not read your whole skill on every turn. It reads only the name and description. When a task matches that description, and only then, it opens the full body. That is why a skill can cost almost nothing in context until the moment it is actually useful — and why you can install a dozen of them without drowning the model.

The idea most founders miss#

Here is the part that separates a skill that works from a folder that gathers dust: the description is the only text that is always in context. It is not documentation. It is the trigger. The model reads it to decide whether to open the skill at all.

A description like "PDF utilities" will lose every race. A description like "Use when converting a customer's Salesforce CSV export into our monthly board-deck revenue format" names the concrete situation, so the model can actually match it. Write the description for the moment of use, not the category. Most broken skills aren't broken — they simply never fire, because nothing ever matched a vague first line.

Build once, run everywhere — that's the real leverage#

Skills follow an open standard, which is the strategic point for anyone building a business, not just a demo. The same folder runs unchanged across Claude.ai, Claude Code, the Claude API, and the Claude Agent SDK. You are not authoring a prompt locked inside one product; you are authoring a portable capability you can move with you. For a solo founder, that means the "how we do things" you encode today survives your next tooling decision. We unpacked why that portability matters in Agent Skills as an open standard.

Skill, MCP server, or subagent?#

The course's most useful clarification is also its least glamorous: skills, MCP servers, and subagents are not competitors, and reaching for the wrong one is the most common expensive mistake.

Most teams over-build an MCP server for a problem a twenty-line SKILL.md would solve. If your agent needs to know something or follow your procedure, that's a skill. If it needs to reach something external, that's MCP. We drew the full decision line in Agent Skill or MCP Server: the 2026 build decision and compared the three primitives head to head in Agent Skills vs Subagents vs Tools.

Do this before Friday#

  1. Take one prompt you paste more than once a week and turn it into a skill — our step-by-step on that exact move is the fast path.
  2. Rewrite its description to name the concrete trigger, not the category. Test whether it fires.
  3. Keep the body lean and push detail into a reference.md the skill loads only when needed — the progressive-disclosure structure is the difference between a skill that scales and one that bloats every turn.

The course is free and short, and the format is genuinely simple. The discipline it asks for is not in the folder structure — it's in that one description line. Get that right and everything downstream is just markdown.