Claude Skills vs OpenAI: Why Developers Can't Stop Arguing About Folders

Anthropic announced Claude Skills in October 2025. Within two days, developers on Reddit were calling it bigger than MCP. OpenAI had nothing to say about it. They were busy with their own stuff.​

I saw the announcement on a Thursday morning. Scrolled past it twice. Thought it was another feature no one would use. Then i watched a dev on Reddit claim he built an indexing flow in 10 minutes. That got my attention.​

Skills aren't plugins. They're not Custom GPTs either. They're folders. Just folders with instructions and files that Claude reads when it needs to. You don't activate them. You don't call them. Claude just knows when to use them.​

The real noise started when people compared it to OpenAI's offerings. Custom GPTs have been around since November 2023. ChatGPT Projects came later. Both do similar things but feel completely different. And that's where it gets interesting.​

What skills actually are

Skills package three things: instructions, code snippets, and reference documents. That's it. You drop them in a folder. Claude accesses them when relevant.​

They work across Claude's web app, Claude Code, and the API. Not just one place. All three. You can combine them for multi-step workflows.​

**Here's what broke my understanding: each skill uses only a few dozen tokens until Claude needs it.**​

Then it loads. This makes them way more efficient than other approaches. I thought everything loaded upfront. Nope.

One developer wrote about this on Reddit. He said skills feel less like installing software and more like giving Claude new thinking patterns. That stuck with me.​

How openai does it differently

OpenAI has Custom GPTs. You define behavior, upload documents, get consistent performance. It's structured. Repeatable.​

Then there's ChatGPT Projects. Similar idea but with organized memory. It stores your instructions and outputs. Good for long-term work.​

The difference is subtle but important. OpenAI's tools are workspaces. You go to them. Claude's skills are capabilities. They come to you during normal conversations.​

Custom GPTs integrate image generation with DALL-E, code interpretation, data analysis. More features out of the box. Claude keeps it simpler.​

What people actually said

Reddit lit up fast. Some loved it. Some didn't.

One user said skills transformed their workflow. Team members could now use CLI tools with plain English. No more teaching coworkers how to navigate the UI.​

Another developer built TypeScript skills that evaluate types through the compiler API. Got specific fast.​

But here's the flip side. Multiple people complained that Claude doesn't follow instructions. One comment got three upvotes: "The primary issue seems to be Claude's failure to adhere to given instructions or utilize the tools it's meant to use."​

That's not small. If the AI ignores your skill, the skill is useless.

Why token efficiency matters

Most people miss this part. Skills only load when needed. That's huge for cost and speed.​

I used to think AI features loaded everything upfront. Context windows fill up fast that way. Gets expensive.

With skills, you can have dozens ready. Claude pulls what it needs. Leaves the rest alone.

This also makes sharing easier. Someone builds a research skill, you grab it, tweak it. Not ideas. Actual logic.​

A developer on Reddit tracked what people built. Community skills popped up everywhere. Rube MCP Connector linked Claude to 500 apps through one server. Saved hours of auth setup.​

The governance thing no one talks about

Claude Code has an "ask-before-act" approach for state-changing operations. File writes, shell commands. It asks first.​

This sounds annoying. It's not. It's the thing that keeps you from accidentally deleting stuff.

OpenAI's Agent Platform leans visual-first. Workflow builders with evaluators and multi-tool connectors. More dashboards. More controls. More to learn.​

Which is better depends on your team. If you want clear ownership and versioning for each capability, Claude's "one package equals one skill" model works. If you're building complex agent graphs with measurement built in, OpenAI wins.​

The weird part about naming things

Why call them skills? Why not plugins or extensions or tools?

Anthropic chose that word carefully. Skills imply behavior. Something learned. Something Claude gets good at.

Plugins sound technical. External. Something you bolt on.

But skills sound internal. Like Claude learned to do your job. Which is kind of what's happening.

The naming makes people treat them differently. I've seen teams share "skills" more freely than they'd share "custom integrations." Psychological thing maybe.

When you shouldn't bother

Most people don't need this. One Reddit user asked if anyone else just describes the task and asks Claude for a plan. No skills, no plugins, no overhead.​

That works fine for one-off tasks. Skills make sense when you repeat things. When you need the same logic every week.

If you're a solo developer doing varied work, skip it. The setup time isn't worth it.

Also skip it if your team won't maintain skills. They go stale fast. Instructions get outdated. Files get replaced. Then Claude uses old information and gives you wrong answers.

Small teams without dedicated AI workflow people? Probably overkill.

Where this goes next

OpenAI's Custom GPTs have been around since late 2023. Claude Skills launched October 2024. OpenAI's had more time to mature.​

But Anthropic's approach feels cleaner. Fewer features. More focus.

The question is whether simplicity wins. Or if OpenAI's feature-rich approach pulls ahead.

Right now developers use both. Claude for some workflows, OpenAI for others. No one's picking sides yet.​

The companies aren't saying much about each other. They're just building. Anthropic rolled out skills. OpenAI's pushing agent platforms and o3 models.​

I keep thinking about that developer who built something in 10 minutes. Ten minutes to set up a flow that would've taken hours before. That's the thing that sticks. Not the features. Not the comparisons. Just the speed of getting real work done.​