Updated May 28, 2026
Stop Prompting, Start Skilling: How to Build AI Skills on Every Platform
Agents get the hype, but skills are what actually ship. A practical guide to building reusable AI skill files in Claude Code, ChatGPT, Copilot, and Gemini.
Agents are sexy. Skills are shipping. Here’s how to build them on every platform.
Everyone wants to talk about agents.
Autonomous AI that runs your business while you sleep. Agents that write your emails, manage your campaigns, review your SEO. The LinkedIn drivel is constant: “I replaced my marketing team with 7 AI agents.”
It’s sexy. It sells. And it’s mostly theater.
Here’s what I’m seeing in the MarketingAlec community. The people who go all-in on agents spend weeks configuring, debugging, and babysitting brittle workflows. The people who build skills ship every day.
Shipping vs. Sexy.
I run and test 30+ AI agents. They’re part of my workflow. But the thing that actually saves me hours every week? A boring set of markdown files that tell Claude exactly how to do repeatable tasks. No autonomy. No multi-agent orchestration. Just: here’s the process, follow it, give me the output. KISS.
Those instruction files are called skills. Right now they’re the single highest-leverage thing you can put in place. They’re also the foundation under the bigger AI marketing skills playbook I keep pointing students toward.
And I keep watching those students hit the same wall: how do you actually build these things? Not the concept. They get the concept. The execution.
Where do I put the file? How detailed should it be? Does this work on ChatGPT, or only Claude?
Let’s get you leveled up.
Video: STOP Prompting, START SKILLING (MarketingAlec on YouTube)
The Treadmill Problem
Most people use AI the same way every session. Open a chat window, ask for what you need, get the output, close the window. Tomorrow, do it all again.
That’s like onboarding a new hire every Monday morning. You explain the brand voice. You describe the project. You list your preferences. The AI nails it. Then the session ends, and all of that context evaporates.
Next time? Start over.
This is the cold start problem, and it’s the number one frustration I see with students. The setup and maintenance cost is the hidden bottleneck in every AI productivity tool. I feel for all the n8n fanboys now babysitting a hundred fragile workflows. You install the tool, run a few generic prompts, get a me-too output, and quietly go back to the old way of doing things.
The fix isn’t a better prompt. It’s a saved instruction set that loads automatically, so the AI already knows who you are, what you’re working on, and how you like things done before you type a single word.
That’s an AI skill.
What a Skill Actually Is
Strip away the branding and a skill is three things in one file:
1. Context. Who you are, what your project is, the resource docs, the conventions that matter.
2. Process. The step-by-step workflow the AI should follow.
3. Guardrails. What NOT to do, the quality checks, the output format.
That’s it. No magic. A markdown document.
I have a skill called “meeting processor.” In plain English, here’s what it does:
- Reads raw meeting notes from a folder
- Pulls out every action item and tags it by owner
- Creates tasks in my task system with the right priority and due date
- Files the meeting notes into the correct project folder
- Logs what it did in my daily note
I wrote it once. It runs identically every time. Last week it processed 24 meetings. This week it’ll do it again. Same quality, same speed, no prompting.
Skills encode structure. The whole point of AI is to compress time. Prompts are a conversation. Structure scales. Random conversations don’t.
Why Now?
Three things happened this spring that pulled skills out of the developer niche and into the mainstream conversation.
Garry Tan went viral. The Y Combinator CEO open-sourced his Claude Code setup: 13 skill files that simulate an entire engineering org. Product manager persona, spec writer, task planner, coder, reviewer. All of it defined as reusable instruction files. More than 20,000 GitHub stars. TechCrunch wrote it up. He claims 100 PRs merged in 7 days using the system.
ChatGPT launched Skills. In early March, OpenAI rolled out a Skills feature for Business and Enterprise users: reusable workflows that ChatGPT auto-applies based on context. This isn’t Custom Instructions. It’s a formalized system for saved process definitions. The name is not a coincidence.
The open standard showed up. The SKILL.md file format, born in Claude Code, is now supported by 16+ tools: Cursor, GitHub Copilot, Gemini CLI, Windsurf, Codex, and more. Write once, run everywhere. Meanwhile AGENTS.md is becoming the cross-platform standard for agent instructions, now backed by the Linux Foundation.
Skills are how we ship. Every single day.
The Cross-Platform Guide
This is the part nobody else is writing. Every blog post, every tutorial, every bootcamp treats skills like they only exist in Claude Code. They don’t. The concept travels.
Claude Code: The Gold Standard
Claude Code has the most mature skill system going. Your master instruction file lives at CLAUDE.md in your project root. Skills live in .claude/skills/ as packaged folders, each with its own instruction file, scripts, and reference docs.
The part that makes it sing: skills auto-trigger based on what you’re doing. You don’t call them. Claude reads the task, scans the skill descriptions, and fires the right one.
Here’s a starter CLAUDE.md you can copy:
# Project Instructions
## About
[Your project/business in 2-3 sentences]
## Voice & Style
[How you want output to sound]
## Conventions
- [Format preference 1]
- [Format preference 2]
- [Tool/platform specifics]
## Common Tasks
- When I say "draft," follow [process]
- When I say "review," check for [criteria]Ten lines. Five minutes to write. And it’ll make every session better from here on out.
ChatGPT: Skills Are Here Now
OpenAI launched Skills in ChatGPT this spring. This isn’t Custom Instructions wearing a new name. It’s a real skill system.
A ChatGPT Skill bundles instructions, examples, and executable code into a reusable package that kicks in automatically when it’s relevant, across web, mobile, and API. Sound familiar? It should. Same concept as Claude Code’s SKILL.md, just built into ChatGPT.
Building one is easy. Every ChatGPT account now ships with a skill-creator baked in. Ask ChatGPT to “create a skill for [your workflow]” and it scaffolds the whole thing: instructions, trigger conditions, output format. Then you can share skills with teammates or publish them to your workspace library.
The catch: it’s currently rolling out to Business and Enterprise users only. On a Plus or free plan, you can fake it with Projects (workspace-scoped instructions plus file uploads) or Custom GPTs (instruction packages with knowledge files and tools).
One real limitation. ChatGPT Skills live inside OpenAI’s UI. You can’t version-control them, export them to a file, or port them to another tool. The portability isn’t there yet.
GitHub Copilot & Microsoft
If you’re in the Microsoft world, it gets more constrained.
GitHub Copilot uses Skills.md, reading a .github/copilot-instructions.md file. You can push further with .github/instructions/*.instructions.md files scoped to specific file paths using applyTo globs.
The headline here: Copilot now natively reads CLAUDE.md, GEMINI.md, and AGENTS.md files. One instruction file, multiple tools. Write it once.
Where it stops: today you cannot just drop SKILL.md into Microsoft 365 Copilot or Copilot Studio and expect it to work the way it does in GitHub Copilot. You’d have to wrap that logic into an Agent Framework or Bot Framework skill or plugin first.
Google Gemini
Gemini CLI runs a GEMINI.md file hierarchy: global instructions in ~/.gemini/GEMINI.md, project-level ones in your working directory, and automatic discovery as Gemini moves through your files. It also supports a SKILL.md system that’s nearly identical to Claude Code’s format.
The thing only Gemini does: @file.md import syntax. You can split your instructions across files and pull them in where you need them. It’s the most flexible instruction architecture of any platform I’ve used.
On Gemini web, the move is Custom Gems. Each Gem is a reusable instruction preset with generous character limits. One Gem per workflow type.
The Write-Once Insight
A single SKILL.md file now works across Claude Code, Gemini CLI, Cursor, GitHub Copilot, and ChatGPT. The format is turning into a cross-platform standard right in front of us.
You’re not locked into any vendor. You’re building a portable skill set that comes with you.
| Platform | Where to Save | Version Control? | Limit |
|---|---|---|---|
| Claude Code | .claude/skills/SKILL.md | Yes (git) | None |
| ChatGPT | Skills (Biz/Enterprise) or Projects (Plus) | No | Generous (Skills) |
| GitHub Copilot | .github/copilot-instructions.md | Yes (git) | ~1,000 lines |
| Gemini CLI | GEMINI.md or SKILL.md | Yes (git) | None |
| Gemini Web | Custom Gems | No | Generous |
| Copilot Studio | Declarative Agent manifest | Via JSON | 8,000 chars |
Your First Skill in 15 Minutes
Don’t overthink this. Pick the task you explain to AI most often and save those instructions.
Step 1: Find your repeat task. Meeting notes, blog drafts, email replies, data analysis, social posts. The one where you type the same context every single session.
Step 2: Write three sections. Context, Process, Guardrails.
Step 3: Save it where your tool can find it.
Step 4: Test it three times. If the output holds across all three runs, you’ve got gold.
Step 5: Iterate. Add a guardrail for whatever went sideways. Skills get better the same way people do, through specific feedback.
Here’s a universal template that works on any platform:
## [Skill Name]
### Context
You are helping [role] with [domain].
My preferred style: [voice/format].
Key context: [project details, brand info, conventions].
### Process
When I ask you to [trigger task]:
1. [First step]
2. [Second step]
3. [Third step]
4. [Output step]
### Guardrails
- Never [common mistake the AI makes]
- Always [quality check]
- Output format: [what the final deliverable looks like]Agents are sexy. Skills are shipping. Build one this week. And if you want the wider map of where this fits, the AI marketing hub lays out the whole function.
Build your first skill, then never start cold again
One email a week, for marketers who’d rather ship than fiddle. The skills I’m actually running, the three-section files behind them, and what each one buys me back once it’s wired in. If “stop prompting, start skilling” landed, the build notes land in your inbox.