How a Skill Loads
Claude can carry a lot of skills, but it doesn't read them all. Think of your skills as a filing cabinet. At startup Claude reads only the labels: every skill's name and description. When a task matches a label, that one drawer opens and its SKILL.md instructions load. Scripts, references and assets come out only when…
How a Skill Loads
Claude can carry a lot of skills, but it doesn't read them all. Think of your skills as a filing cabinet. At startup Claude reads only the labels: every skill's name and description. When a task matches a label, that one drawer opens and its SKILL.md instructions load. Scripts, references and assets come out only when the job needs them. That's progressive disclosure, and it's why the description is the part of a skill that does the real work.
Liam, in for Bear.
Topic: Claude · Agent Skills
Chapters
0:00 The question, and three terms 0:25 The cabinet, and one drawer 0:51 At startup, only the labels 1:01 A task opens one drawer 1:11 One file at a time; the rest stays shut 1:32 Write the label that opens the drawer 1:48 Your turn
YOUR TURN
Paste this into Claude:
Here's a task I repeat every week: [describe it in two sentences]. Draft a SKILL.md for it. Give it a short lowercase name with hyphens, and a description that says what it does and exactly when to use it, in the words I'd actually type. Then write the steps below, and tell me what belongs in a separate reference file or script.
Then check two things yourself:
- Write three ways you'd ask for it. Does the description catch all three?
- Name one request that should not open it. Does it stay out of the way?
Sources
- Anthropic, Agent Skills repository: github.com/anthropics/skills
- Agent Skills specification: agentskills.io/specification
@NikBearBrown · Part of the Claude & Agentic AI series.
▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU
What this video covers
- 0:00 — The question, and three terms
- 0:25 — The cabinet, and one drawer
- 0:51 — At startup, only the labels
- 1:01 — A task opens one drawer
- 1:11 — One file at a time; the rest stays shut
- 1:32 — Write the label that opens the drawer
- 1:48 — Your turn
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
The question, and three terms
[0:00] Ola, this is Liam in for Bear. Claude can carry a lot of skills, but it doesn't read them all. So, don't ask how Claude reads every skill. Ask how it knows which one to open. Three terms: a skill, a folder of instructions, scripts, and resources that Claude loads for a task. Skill.md, it's one required file with a name and description on top and instructions below. And progressive disclosure, more detail loads only as the task needs it.
The cabinet, and one drawer
[0:25] Picture your skills as a filing cabinet. Each drawer is one skill, and on the front of every drawer, there's a label. Pull one drawer out. Up front is the one file every skill needs, skill.md. It's top holds a name and a description of what the skill does and when to use it. That's the label. Below it, the instructions. Behind it, three optional folders: scripts, code Claude can run, references, documents it can read, and assets like templates and images.
At startup, only the labels
[0:51] Here's the trick. When a session starts, Claude doesn't open the cabinet. It reads only the labels, every skill's name and description. That's all that's always in context. Then a task arrives,
A task opens one drawer
[1:02] Fill in this PDF form. Claude checks it against the labels. The PDF skill matches, so only that drawer slides open and its instructions load. Those
One file at a time; the rest stays shut
[1:11] Instructions point to more files, and Claude pulls one only when the job needs it. Filling a form, read the form's guide. A script can even run without its code being loaded. Only its output comes back. Every other drawer stays shut, so you can install many skills, and each one costs little more than its label until a task needs it. That's progressive disclosure.
Write the label that opens the drawer
[1:32] So, the description does the real work. It's the label that decides whether the drawer opens. The specs own bad example is helps with PDFs. Say what the skill does and when to use it in the words people actually type. An Anthropic skill guide says Claude tends to underuse skills, so make it a little pushy.
Your turn
[1:48] Your turn, paste this into Claude. Here's a task I repeat every week. Describe it in two sentences. Draft a skill.md for it. Give it a short lowercase name with hyphens and a description that says what it does and exactly when to use it in the words I'd actually type. Then write the steps below and tell me what belongs in a separate reference file or script. Then check two things yourself. Write three different ways you'd ask for this task. Does the description catch all three? And name one request that should not open it. Does the description stay out of the way? How a skill loads at Nick Bear Brown.
