Section 1
What's a Skill, Really?
Skills are folders. That's it.
Inside each folder is a text file with instructions. When you trigger it, Claude reads those instructions and follows them. You're not writing code. You're writing a recipe, and Claude is the chef.
Why does this matter? Because right now, every time you do a repeated task with Claude, you're re-explaining the same thing from scratch. Your voice. Your format. Your workflow. Skills let you teach Claude once and reuse it forever.
The mental model:
- A skill = a folder
- Inside the folder = a text file with task instructions
- You say a trigger phrase → Claude reads the folder → Claude does the task
Section 2
Install the Skill Creator
You don't have to learn YAML. You don't have to memorize the folder structure. Anthropic ships a skill called Skill Creator that interviews you about the task and writes the file for you.
Same skill, three entry points. Install it in whichever Claude you already use.
Claude.ai (web)
- Go to claude.ai and open Settings, then Capabilities.
- Find Skill Creator in the list of Anthropic skills.
- Toggle it on. That's it.
Now in any conversation, you can say "help me build a skill" and it activates.
Claude Cowork (Desktop)
- Open the Claude Desktop app and start a new conversation.
- Type
/in the message box to open the skill menu. - Browse to Skill Creator, click install.
Every skill you build from here shows up in the same / menu. Claude also auto-detects when one is relevant and loads it for you.
Claude Code (CLI)
- Inside an active Claude Code session, run
/plugins. - Search for skill-creator and install the official Anthropic version.
- The skill lands in
~/.claude/skills/skill-creator/or your project's.claude/skills/.
Any skills you build with it save to the same folder and are auto-loaded in every future Claude Code session.
Heads up: Skills you build in one Claude do not auto-sync to the other two. Build where you work. If you live in Claude Code, build there. If you use Claude.ai in the browser, build there. Most people only need one.
Section 3
Your First Skill in 15 Minutes
Once Skill Creator is installed, building a skill is a conversation. No coding, no YAML, no folder wrangling. You answer a few questions. It writes the file. You save it. Done.
Pick a task you do on repeat
Not your hardest work. The most boring repeatable thing. The task you've explained the same way five times this week. Weekly client report. LinkedIn post from an article. Follow-up email after a call. That one.
Trigger the Skill Creator
Open a fresh conversation and say "I want to build a skill for [task]". The Skill Creator activates and starts the interview.
Answer the 4 phases
Skill Creator runs a 4-phase interview: Identify the task, extract your current process, optimize before building, then write the file. The optimize phase is the one most people skip. Don't. That's where your current process gets upgraded into a better one before it gets encoded forever.
Save the file it writes
In Claude Code, Skill Creator writes directly to ~/.claude/skills/[name]/SKILL.md. In Claude.ai and Cowork, the skill saves to your account automatically. Don't overthink this step.
Test it with a real input
Open a brand new conversation. Use the phrase you'd naturally say to kick off the task. Check the output. If the skill didn't trigger, the description is too vague. If the output is wrong, the instructions need more specifics. Fix and run it again. First skill always takes one round of edits.
That's the full loop. 15 minutes for most skills, start to saved-and-working. You now own a piece of automation that runs the same way every time, forever.
Section 4
Under the Hood
You don't have to understand this to use Skill Creator. But if you want to edit a skill directly, or write one by hand, here's how they actually work.
The folder structure
Every skill in Claude Code lives in .claude/skills/. Each skill gets its own subfolder with at minimum one .md file inside.
.claude/
skills/
weekly-report/
SKILL.md ← the main skill file
social-post/
SKILL.md
brand-voice.md ← linked reference file
client-summary/
SKILL.md
template.mdThe 3-level loading system
Claude doesn't read everything at once. It loads skill context in three layers, and understanding this is what separates good skills from great ones.
Frontmatter
Always loaded. The YAML block at the top of your skill file. Contains the trigger description, name, and tags. Claude reads this every conversation to know when to activate your skill.
Keep it under 200 tokens. Every conversation pays this cost.
Skill Body
Loaded when triggered. The full instructions, steps, examples, and formatting guidance. This is where most of your content goes.
One skill, one job. Don't pack multiple workflows into a single file.
Linked Files
Loaded on demand. If your skill references a brand guide, template, or example library, Claude only loads those files when it actually needs them. Keeps the context window light.
Use relative paths: ./brand-voice.md
Section 5
Writing Descriptions That Trigger
The frontmatter description is the single most important line in your skill file. Write it wrong and the skill either never triggers or triggers at the wrong time. Skill Creator writes a decent first version. These are the rules for making it great.
---
name: weekly-report
description: "Generate weekly client status report from project notes using the standard 3-section format"
---Too vague. Won't trigger.
"Create report"
Specific. Triggers right.
"Generate weekly client status report from project notes"
Too broad. Overlaps.
"Help with writing tasks"
Scoped. Fires at the right moment.
"Write LinkedIn post from newsletter content using brand voice"
Instructions that work vs. instructions that don't
The body of your skill is where the task gets done. Claude follows specific, action-oriented language much better than vague guidance. Rewrite anything abstract into something concrete.
Vague
"Make the report professional and clear"
Actionable
"Use H2 headers for each section. Keep bullets to one sentence. No em dashes."
Vague
"Include the important information"
Actionable
"Include: status (one word), blockers (if any), next action (one sentence)"
Section 6
The 5 Use Cases That Actually Matter
Skills aren't useful for everything. They're useful for 5 specific kinds of tasks. Ranked by raw payoff, most-to-least. Every example below is a real skill I built and run daily.
Multi-step workflow automation
Processes with 4+ steps that run the same order every time. Onboarding flows. Publishing pipelines. Review checklists. One skill runs the whole thing.
Real example
lead-magnet-launcher
Launching a new free guide used to take me 2 hours: write the guide page, write landing copy, write the delivery email, add a card to the resources page, build the landing page, push to GitHub. Now it's one skill. One prompt. 6 outputs in 15 minutes.
This is the "holy sh*t" moment of skills. Most people don't realize one skill can orchestrate a full pipeline.
Content production with format consistency
Anything you produce in a specific format, voice, or structure. Social posts. Newsletters. Reports. Scripts. The skill enforces the shape so you never re-explain it.
Real examples
short-form-script-writer
Writes TikTok and Reels scripts using the exact hook patterns, CTA placements, and format from the scripts that pulled 1.5% follower conversion. I never open a blank page. I say "script this idea" and the skill writes it in my voice.
instagram-carousel
Writes slide copy and renders full 1080×1350 carousel PNGs via Playwright. From topic to upload-ready images in one step.
Format never drifts. Voice never drifts. You never explain your style again.
MCP enhancement
Think of it like this. MCP gives Claude the kitchen. Skills give Claude the recipes. Most people install MCPs and stop there, then wonder why Claude still feels generic. Skills are what make your MCPs useful for your specific work.
Real example
tiktok-intelligence
I have a Supabase database of competitor TikTok performance data. Alone, the Supabase MCP just gives Claude raw access. The skill wraps it and tells Claude exactly which tables to query, how to filter by engagement, and how to format the output. Before I write any script, I run this skill and Claude tells me which hooks are working right now.
The most slept-on use case. Your MCPs are only as good as the skills telling Claude how to use them.
Tiny utility skills
Not every skill has to be a workflow. Some are 20 lines of instructions that just do one annoying cleanup task, well, every time. These are perfect first skills. Low risk, instant payoff, and they build the habit of skill-ifying.
Real example
em-dash-destroyer
AI writing is littered with em dashes. I hate them. This skill runs after any piece of writing, strips every em dash, and replaces it with better punctuation. 20 lines. Runs in seconds. Saves me 5 minutes of find-and-replace on every draft.
Build one of these first. You'll hate going back to doing it by hand.
Orchestration and meta-skills
Once you have 5 to 10 skills, a new problem appears: which one do I run, and when? Meta-skills solve that. They take a vague request, figure out what you actually need, and route you to the right skill (or chain of skills).
Real example
orchestrator
When I say something fuzzy like "I need to launch something new this week," the orchestrator asks two or three qualifying questions, then recommends a skill sequence: run positioning-angles, then direct-response-copy, then lead-magnet-launcher. It's the conductor for the rest of the library.
Advanced move. Don't start here. But once your library gets past 10 skills, orchestration compounds everything.
Want the actual .md files for any of these? I break down skills like these every week in my newsletter, so you can drop them into your own setup, modify them, or use them as scaffolding for your own versions.
Section 7
Testing Your Skill
Built your first one? Run these three checks before you depend on it for real work.
1. Does it trigger?
Type the phrase you'd naturally use to start the task. Claude should recognize it and activate the skill. If it doesn't, rewrite the description to match how you'd actually phrase the request.
2. Does it do the right thing?
Run it with a real example and compare the output to what you wanted. Does it ask for the right inputs? Does it use the right format? Are the instructions being followed? Catch gaps now, before the skill gets used 50 times with the same flaw.
3. Is it fast enough?
Skills that load large reference files can feel slow. If you notice lag, move big files to linked references instead of including everything in the main skill body.
Section 8
Common Mistakes
Trying to skill-ify strategic thinking
Skills are for repeatable output tasks with consistent inputs. "Think through whether to launch X" is not a skill, it's a one-off decision. If the task changes shape every time, it's a prompt, not a skill.
Encoding your current process instead of your best one
The Skill Creator's optimize phase exists for a reason. If you skip it, you just locked in your current workflow, bugs and all. Always answer the "what would a better version look like?" question before saving.
Vague trigger descriptions
If your description is too generic, Claude either won't activate the skill or will activate it at the wrong time. Be specific about the task, the input, and the output.
Packing too much into one skill
One skill, one job. A skill that does five things does none of them well. Split complex workflows into multiple skills and chain them manually, or build an orchestrator that routes between them.
Forgetting to test with messy real-world input
Skills always look good with the clean example you had in your head. Test with the ugliest real input you have: partial info, weird formats, edge cases. That's when you find the gaps worth fixing.
Never updating the skill
The first version is a draft. Use it for a week, notice what drifts, update the instructions. Skills compound with iteration. They're not set-and-forget.
Section 9
Start Here
Don't try to build your dream skill first. Pick the most boring, most repeatable task you've done 3+ times this week. Something you'd explain to a new hire in under 5 minutes.
Install the Skill Creator using the section above. Open a new conversation. Say "I want to build a skill for [that boring task]." Let it interview you. Save the file.
Test it once. Iterate. Add the second one tomorrow. That's the whole system. Most people never build their first. You'll be further than 90% of Claude users by Friday.
Manual quick start (Claude Code)
$ mkdir -p .claude/skills/my-first-skillThen create SKILL.md inside that folder. Or just use Skill Creator and skip the manual work.
Quick Reference
Skill Builder Cheat Sheet
Install Skill Creator (web)
claude.ai → Settings → Capabilities
Install Skill Creator (desktop)
Type / in Cowork → browse skills
Install Skill Creator (code)
/plugins → search skill-creator
Skill location (Claude Code)
.claude/skills/[name]/SKILL.md
Trigger
Frontmatter description field, be specific
Loading order
Frontmatter → body → linked files
Frontmatter limit
Under 200 tokens (every conversation pays)
Reference files
Link with ./filename.md relative paths
Best use case #1
Multi-step workflow automation
Best use case #2
Content production with format consistency
Best use case #3
MCP enhancement (recipes for your tools)
Best use case #4
Tiny utility skills (start here)
Best use case #5
Orchestration and meta-skills
Good description
"Generate weekly client report from notes"
Bad description
"Help with writing" (too vague)
Golden rule
One skill = one job. Split complex workflows.
First test
Run it with the messiest real input you have
Work with Me
Need AI to actually work for your business?
I help businesses cut through the AI hype and build the workflows, automations, and systems that actually move the needle. Direct, hands-on, no fluff.
Work with me