Claude Skills
Build your first Claude skill in five minutes, no code.
A skill is a folder Claude reads every time. Inside it: your taste, your rules, your steps. You build it once, Claude fires it automatically, and you stop pasting the instructions into every chat.
4
Steps
5 min
Build time
0
Lines of code
A skill is just a folder. That's the whole trick.
Most people prompt Claude the same way every time. The same brand rules, the same tone, the same checklist, copy-pasted into every chat. Skills move that knowledge out of the prompt and into the model.
A skill is a folder with one file inside it called SKILL.md. Inside that file: instructions Claude reads every time the skill is loaded. Once it is installed, you stop telling Claude what your voice is. You stop pasting your design system. You stop re-explaining your client's brand. The skill does that work silently, on every prompt.
The thing that makes this trickier than it looks is the name and description at the top of the file. That is the trigger surface. Get it wrong, the skill never fires. Get it right, the skill fires when it should without you naming it.
The shift: your knowledge moves from the prompt to the model. You write it once, Claude runs it forever.
STEP 01 · MAKE
Open the create flow
Let Claude interview you.
Writing a SKILL.md from scratch sounds like documentation work. Most people open a blank file, type three lines, get bored, close it. The skill never gets made.
The shortcut: don't write it. Let Claude write it. The Customize panel has a "Create with Claude" option that runs a short interview. You describe the skill in plain English, Claude asks four or five questions, the .skill file drops into your downloads folder.
The interview catches what you would have forgotten. When should the skill auto-trigger? What edge cases? What output format? These are the parts you skip if writing from scratch, and they are the parts that decide whether the skill actually works
Example: GENERATE A BRAND-VOICE SKILL
You tell Claude: "Make me a skill that rewrites any draft in my brand voice. I'll answer your questions."
STRENGTHS
- Skips the blank-file paralysis.
- Claude catches edge cases you'd miss.
- File is ready to upload in two minutes.
LIMITS
- The interview is only as good as your descriptions. Vague answers, vague skill.
- You still need to test what auto-triggers and what doesn't.
- Doesn't help if your skill needs custom code logic. Very few do.
Use this when: you're building your first skill, or any skill where the logic lives in instructions, not code.
STEP 02 · INSTRUCT
The SKILL.md anatomy.
Name and description do the heavy lifting.
Most people obsess about the instructions. The instructions matter less than the metadata.
The shortcut: the frontmatter at the top of every SKILL.md, the name and description, is the trigger surface. Claude scans descriptions across every installed skill and decides which to load. Get the description wrong and the skill never fires, no matter how good the instructions below it are.
A description that works names the trigger condition first, the function second, the edge cases last. Use this skill when [trigger]. Covers [function]. Also trigger when [edge cases]. Specific verbs over abstract ones. Named outputs over described ones.
Example: A SKILL.MD THAT TRIGGERS RELIABLY
Claude writes the file during the interview. You review and edit only the description.
STRENGTHS
- Name and description is the only thing that decides triggering.
- "Use this skill when X" is the format Claude is trained on.
- Edge case clauses catch the 20% you'd lose otherwise.
LIMITS
- A description too narrow never fires.
- A description too broad fires on the wrong prompts.
- Editing the description is the iteration loop. First version rarely lands.
Use this when: the auto-trigger isn't firing reliably, or the skill is firing on prompts where it shouldn't.
STEP 03 · UPLOAD
Drop the file back in.
Customize, Skills, Upload.
Claude generated a skill file. It is sitting in your downloads folder. To use it, you upload it back through the same Customize panel where you started.
The shortcut: Customize, Skills, hit the plus icon, choose Upload a skill, drag the .skill file in. The skill appears in the list, ready to test.
The list view also shows try, edit, replace, download, and uninstall options. Most operators forget the edit option exists. Skills are not write-once. The first version is rarely the version that ships.
Example: UPLOAD AND VERIFY
You drag brand-voice.skill into the upload area.
STRENGTHS
- One drag installs the skill.
- Available immediately across every chat.
- Edit-in-place means you iterate without re-uploading.
LIMITS
- No version history. Replace overwrites the previous file.
- No "test against last week's chats". You run new prompts.
- Skills don't transfer between accounts. Team sharing needs re-upload.
Use this when: the .skill file is generated and you're ready to test it on a real prompt.
STEP 03 · TRIGGER
Two ways to fire it.
Auto when it fits, manual when you mean it.
Two trigger surfaces: auto and manual. Auto fires when Claude reads the description and decides the skill is relevant to what you're doing. Paste a draft into chat and the brand-voice skill auto-loads.
The shortcut: type /skill-name to fire it manually. This bypasses Claude's interpretation. Use it when you want a specific skill and don't want to gamble on whether the auto-trigger reads your intent right.
The mark of a well-written description is that the skill auto-fires when it should, without you naming it. The first version usually doesn't. Watch what triggers and what doesn't over a few real prompts, then edit the description. That is the loop.
Example: A SKILL FIRING ON A PASTED DRAFT
You paste a blog draft and type "polish this for me".
STRENGTHS
- Auto-trigger means the skill works without you remembering to call it.
- Manual /skill-name is the override for ambiguous prompts.
- Multiple skills can load in one chat. They layer.
LIMITS
- Skills with overlapping descriptions fight. Claude picks one.
- Auto-trigger is a probability call. Ambiguous prompts miss the right skill.
- You won't always see which skill fired unless you check.
Use this when: the skill is uploaded and you want to know whether the description is doing its job. Watch the trigger detection, edit the description, re-test.
The four steps at a glance
How to start.
Pick the prompt you've pasted into Claude three or more times this month. That's your first skill. The repetition is the signal. You've already written it, you just keep re-writing it.
Open Customize, hit Create with Claude, describe what you keep doing. Answer Claude's questions in plain English. Don't worry about the SKILL.md syntax. Claude writes the file.
Upload it, try it once, edit the description if the auto-trigger doesn't land. That's the whole loop. After the second skill, the loop is muscle memory.
The first skill is the slowest. Every one after teaches you the shape of the next