What the Hell Are Claude Skills, CustomGPTs, Gemini Gems?
Different names, same idea?
Every big AI company lets you build your own little assistant.
OpenAI has CustomGPTs. Google has Gems. Anthropic has Skills. Microsoft and Mistral have their own versions of the same thing. Or do they? Are these the same thing or not?
They sound like five different products. Or they’re one product wearing five outfits.
Let me explain what these actually are. Because the marketing pages won’t.
The same trick
None of these are new AI models. They’re all the same model you already use, with a costume and a briefing.
A CustomGPT is still ChatGPT. A Gem is still Gemini. A Skill is still Claude. You didn’t build a smarter AI. You wrapped the existing one in (smarter) instructions.
The wrapper usually has three parts:
Instructions. A standing prompt that tells the model who to be and how to act. “You’re my newsletter editor. Cut fluff. Keep my voice.”
Knowledge. Files you upload so it can answer from your stuff, not just its training. Your style guide, your past work, a product manual.
Tools. Sometimes. The ability to actually do something, like call an API or run code.
So: instructions tell it how to act. Knowledge tells it what to know. Tools let it act.
Most of these wrappers stop after the first two. The third one is where they split apart, and I’ll get to that.
CustomGPTs
OpenAI started this race in late 2023.
You give a GPT some instructions, upload a few files as knowledge, switch on capabilities like web browsing or image generation, and you can wire it to an external API with something called Actions. Then you share it with a link.
A simple example. Say you run a shop. You could build a GPT that knows your return policy, answers in your tone, and checks live order status through your store’s API. Set it up once, then everyone just chats with it.
OpenAI built a whole GPT Store around this. You’d build a clever GPT, people would use it, and you’d get paid. A little app store for prompts.
There are thousands of GPTs now and almost nobody finds them. I wrote about the GPT Store when it launched.
As a tool for yourself, a CustomGPT is fine. You can shape it to your liking and it remembers the setup.
Gems
Google’s version is the simplest of the three.
A Gem is a saved set of instructions for Gemini, with optional file uploads for context. Google ships some premade ones (a coding partner, a writing editor, a brainstorm buddy) and you can make your own.
Think of it as a chatbot with a job description pinned to it. A Gem that always edits to your style rules. A Gem that knows your coding stack so you stop re-explaining it. You set the briefing once, then talk to it.
That’s basically it. No store. No API actions for most people. It lives in the Gemini app and inside Google Workspace, so a Gem can lean on your Docs and Drive if you let it.
If a CustomGPT is a saved prompt with extras, a Gem is a saved prompt with fewer extras. Lighter. Less to configure. Also less it can do.
Skills
Anthropic’s Skills are the one that’s a bit different. They’re also the newest, from late 2025.
A Skill is a file, or a set of files, or a folder. Inside sits a set of instructions, and next to that you can put scripts, code, and resources the model can use.
Two things make them stand out.
First, Claude loads a Skill only when the task calls for it. You don’t pick it from a menu. Claude reads a short description of each Skill you have, and pulls in the full thing the moment the work matches. So you can have a stack of them sitting there and the right one shows up on its own.
Second, the tools part. A Skill can carry real, runnable code, and Claude can execute it. There are Skills that build an actual Word document, fill a PDF, or run a script against your data. Not describe the file. Make the file. So a Skill doesn’t just say how to do a thing. It can do the thing.
They work across Claude’s apps, the API, and Claude Code, the coding tool I use every day.
In December 2025 Anthropic published the Skill format as an open standard. Which means a Skill you build isn’t trapped inside Claude. Other tools that adopt the standard, like Cursor, can run it too. That’s the opposite of how a CustomGPT or a Gem works.
A CustomGPT and a Gem mostly talk. A Skill can act. And it can travel.
The rest
Everyone else has a flavor of this too.
Microsoft has Copilot agents and Copilot Studio, aimed at companies wiring AI into their internal systems. Mistral’s Le Chat has agents.
And then there are Projects in ChatGPT and Claude. Custom instructions buried in your settings. Those are versions of the same trick. A standing instruction the model reads every time, so you don’t repeat yourself. With the chat memories.
Talk or do
If you want one way to sort all of this, here it is.
Does the wrapper just talk, or can it run something.
Most of them just talk. A CustomGPT with knowledge files is a better-briefed chatbot. A Gem is a better-briefed chatbot. Useful, but you could get 90% of the way there by pasting a good prompt yourself.
The ones that can run code, like Skills, or a CustomGPT hooked to a real API through Actions, are a different category. They reach out and change something in the real world. That’s the part worth caring about.
The second question is whether you’re stuck. A Gem only lives in Gemini. A CustomGPT only lives in ChatGPT. A Skill, now that it’s an open standard, can move. If you’re going to invest hours building one of these, that matters more than it sounds.
Everything else is presets.
Downsides
Lock-in first, because for most of them it’s the big one.
Your CustomGPT only works in ChatGPT. Your Gem only works in Gemini. Build your workflow on one of those and you’ve quietly chosen that company for as long as you use it. Skills are the exception, since the format is open now, but they’re still the new kid and the ecosystem around them is young.
Second, a wrapper is only as good as the model under it. A clever costume on a weak model is still a weak model. You’re not buying intelligence. You’re buying a configuration.
Third, the knowledge goes stale. You upload your files once and forget them, and six months later the assistant is confidently working from an old version. Nobody sends you a reminder.
The Bottom Line
These are not five inventions. They’re one idea with five names.
Instructions, maybe some files, maybe the ability to actually do something. Pick a name based on which AI you already pay for, not because one of them is secretly a different species.
The only split that means anything is whether the thing can run code or only chat, and whether it can leave the building it was born in. That’s what I’d ask before building anything on top of one of them.
The rest is branding.
A writer is nothing without a reader. If you found this helpful, consider becoming my dear email friend. Nothing would make me happier.