You can now hand an AI a folder of instructions and have it behave like a specialist. Two ways to do that dominate right now: Claude skills (the SKILL.md format Anthropic introduced) and OpenAI's custom GPTs. They get lumped together constantly, and they shouldn't be, because they solve different problems and fail in different ways.
I've built and tested both formats for months. Here is the honest comparison, including the cases where a custom GPT is the right call.
The short answer
A custom GPT is a hosted chatbot with a personality and some reference files, living inside ChatGPT. A Claude skill is a portable instruction file (or folder of files) that any capable AI can load and execute like a procedure. If you want to share a link with your audience, build a GPT. If you want a repeatable process that produces consistent work output, whether that's pricing a project, running a negotiation prep, or qualifying a lead list, skills win, and it isn't close.
What a custom GPT actually is
When you create a custom GPT, you fill in a system prompt, upload up to 20 knowledge files, toggle a few capabilities (browsing, image generation, code interpreter), and publish. OpenAI hosts it. Anyone with a ChatGPT account can use it from a link.
That hosting is the whole appeal and the whole limitation. Your GPT lives inside ChatGPT, runs on whatever model OpenAI routes it to, and behaves however that model feels like behaving today. You can't version it properly, you can't run it in Claude or Gemini, and if OpenAI changes the underlying model, your carefully tuned instructions can drift overnight. Builders have complained about exactly this after nearly every major model update.
There's also a depth ceiling. A GPT is one system prompt. You can write a long one, but it's still a single flat instruction set. The model reads it once and then improvises. For a persona ("talk like a pirate chef") that's fine. For a process with steps that must happen in order, it's shaky.
What a Claude skill actually is
A skill is a markdown file, SKILL.md, with a small metadata header and a body of instructions. The header tells the AI when the skill applies. The body tells it exactly how to do the job: steps, rules, output formats, edge cases. Skills can also ship as folders with extra reference files and scripts the AI reads only when needed.
Three properties make this format different in practice.
First, it's portable. A SKILL.md is plain text. It works in Claude, in Claude Code, and with light adaptation in ChatGPT Projects or Gemini Gems. You own the file. Nobody can deprecate it out from under you.
Second, it loads on demand. Claude reads the metadata of every installed skill (a few dozen tokens each) and only pulls the full instructions into context when the task matches. That means you can install thirty skills without drowning the model in irrelevant instructions. A custom GPT carries its entire system prompt into every single message.
Third, and this is the one that matters most for real work: skills can encode sequence. A well built skill says "do A, then B, and do not produce the final answer until C is confirmed." At PromptLeadz we call these checkpoints gates. Our Meta Negotiator skill, for example, refuses to script your counter-offer until it has forced you through leverage analysis and a walk-away number, because a counter-offer without those two things is guessing. A single flat system prompt almost never holds that discipline. The model skips to the answer it thinks you want.
Head to head
| Custom GPT | Claude skill | |
|---|---|---|
| Format | Hosted chatbot config | Portable markdown file(s) |
| Where it runs | ChatGPT only | Claude, Claude Code; adaptable to ChatGPT Projects and Gemini Gems |
| Ownership | Lives on OpenAI's servers | A file you keep |
| Context cost | Full instructions in every message | Loads only when relevant |
| Multi-step process discipline | Weak; model improvises | Strong; steps and gates can be enforced |
| Composability | One GPT at a time | Skills stack and chain |
| Sharing with an audience | Excellent, one link | Send the file, or a zip |
| Discovery / storefront | GPT Store | No central store yet |
| Version control | Manual, awkward | It's a text file; git works |
Where custom GPTs still win
Credit where due. If your goal is distribution to a mass consumer audience, the GPT Store and a shareable link beat emailing people a markdown file. A creator with 100k followers who wants a branded chatbot should build a GPT. Same if your users live in ChatGPT and will never leave it, or if the use case is a light persona rather than a process.
GPTs are also faster to stand up. Ten minutes in the builder and you have something usable. Writing a good skill takes longer because you're documenting a real procedure, and most people have never written their process down at that level of detail. That effort is exactly why the result is better, but it is real effort.
Where skills win
Anything you'd call work product. Pricing and quoting. Sales call preparation. Interview practice with scoring. Contract review checklists. Lead qualification. Anything where the output must be consistent across runs and where skipping a step produces a confident wrong answer.
The stacking matters more than people expect. Because skills load on demand, you can build a library where one orchestrator skill routes to specialists. Describe the task in plain words, the orchestrator classifies it and pulls in the right sub-skill, or chains several in sequence. That architecture simply doesn't exist inside a single custom GPT. It's the difference between hiring one generalist and running a small team with a coordinator.
There's a quieter advantage too: model independence. The last two years should have taught everyone that betting your workflow on one vendor's chat interface is fragile. A skill file survives you switching from Claude to whatever comes next. A custom GPT does not.
A worked example: pricing a freelance project both ways
Abstract comparisons hide the difference, so here's a concrete one. Say you're a freelance designer quoting a brand refresh, and you ask an AI to help you price it.
The custom GPT version: you open a "Freelance Pricing Assistant" GPT from the store. It asks what the project is, you describe it in two sentences, and it produces a tidy quote with three tiers. Looks professional. The problem is what it never asked: your costs, your capacity this quarter, what the client's alternative options cost, whether there are usage rights involved, what happens on revision three. The number is plausible and unanchored. Plausible and unanchored is how freelancers underprice by 40%.
The skill version runs differently because the file forces it to. A pricing skill with gates will not output a number until it has walked scope (what's actually included, what's explicitly excluded), effort (your realistic hours, not your optimistic ones), market anchor (what this work sells for, not what you're used to charging), and risk (revision limits, kill fees, rights). Only then does it price, and the quote comes with the reasoning attached, so when the client pushes back you know exactly which lever moved.
Same model underneath. Same question from you. The difference is that one format lets the AI answer immediately and the other one makes it earn the answer. That gap shows up in every process-shaped task: negotiation prep, discovery calls, hiring interviews, investor Q&A rehearsal.
Common mistakes when writing your first skill
If you do write your own, four failure modes account for most bad skills.
Writing goals instead of procedure. "Help the user price confidently" is a wish. "Ask for the deliverable list. Refuse to continue until you have it" is an instruction. Skills work when they read like a checklist a competent stranger could follow.
Stuffing everything into one file. When a skill tries to cover pricing and proposals and negotiation, the model gets vague on all three. One job per skill. Let an orchestrator route between them; that's the architecture our brains use, ten specialists plus a coordinator rather than one bloated generalist.
No output contract. If you don't specify what the finished product looks like (a table, a script, a number with reasoning), you'll get a different shape every run, and you can't build a workflow on outputs that change shape.
Skipping the trigger metadata. The description header is how the AI decides when the skill applies. Write it vaguely and your skill either fires constantly or never.
Can you use both?
Yes, and the pattern that works is: author once as a skill, deploy everywhere. Keep the SKILL.md as your source of truth. Paste or upload it into a ChatGPT Project when you need it there, drop it into a Gemini Gem, install it natively in Claude. All of our PromptLeadz skills ship this way, one file, three platforms, because the instructions are the asset, not the wrapper.
How to try skills without building one
Two routes. If you want to learn the format, Anthropic's docs cover the SKILL.md spec and you can write a simple one in an afternoon; start with a task you repeat weekly and already do well, and transcribe your own process. The skill will be as good as your documentation of it.
If you'd rather start from a working example, we give away complete skills. Ask us for any SKILL.md and we build it free, or pick up one of the flagship brains like Meta Closer (sales, first call to signature) or Meta Strategist (decision pressure-testing). The Vault is the whole library for $99, lifetime, including everything released later.
FAQ
Are Claude skills the same as Claude Projects?
No. A Project is a workspace with files and custom instructions attached, closer in spirit to a custom GPT. A skill is a self-contained procedure file that loads when a task calls for it. You can put skills inside a Project, which is the best of both.
Do skills work in ChatGPT?
The format is Anthropic's, but the file is plain markdown. Upload it to a ChatGPT Project or paste it as instructions and most well written skills run fine. You lose the automatic on-demand loading, which is a Claude feature, but the procedure itself carries over.
Can a custom GPT do multi-step workflows?
It can attempt them. Without enforced checkpoints it tends to compress steps and jump to conclusions, especially on long tasks. If step order matters to correctness, you want the skill format.
What does a good skill cost?
Free to write yourself. Pre-built professional ones run anywhere from free to a few hundred dollars. Ours are $5 to $99, and the free tier is a complete working file, not a teaser.
Which should a small business owner pick in 2026?
If you're automating your own operations: skills. If you're shipping a branded chatbot to customers inside ChatGPT: a custom GPT. Many businesses end up with both, authored from the same skill files.
PromptLeadz builds framework-based AI skills, single files that install into Claude, ChatGPT, and Gemini and behave like specialists instead of chatbots. Start with a free custom SKILL.md.
Leave a comment: