I use LLMs for a ton of things, but every so often — on some random, unremarkable day — I learn they can do something I never imagined they could. And with agentic AI in the mix, those moments have only gotten stranger. The feature responsible for the latest of mine isn’t the one everybody talks about. It isn’t Claude Code. It’s skills.

Skills are really a “quiet” upgrade. You see, with AI, we’re now at a point where the core models are all good enough. The difference in the experience comes down to the prompts, and the tools available to the model. Claude supports skills, and it can also build them for itself on the spot. The latter is what makes it occasionally feel almost magical. And all it takes for you is to ask it nicely.

“Learn how to”

Ask and ye shall receive

Credit: MUO

When I reviewed OpenClaw, I mentioned a moment that stuck with me. I was chatting with OpenClaw over Telegram and asked whether it could send voice messages. It said no. So I told it, plainly: learn how to send voice messages. It went off, installed the packages and libraries it needed, wrote itself a README on how to operate them, and, two messages after, I had a voice note sitting in my chat. It just did it.

Now I know that capability isn’t unique to OpenClaw. The chatbots we use all are also agents. They have their own sandboxes to run stuff on. There’s a whole concept of skills that lets LLMs do basically anything, and it gets even more powerful when you pair it with MCP servers (we won’t talk about MCP servers here).

Claude does this too, and arguably does it better, because Claude Code and Claude Cowork are baked right into the Claude app. Once you build a skill and add it, it stays there. Where OpenClaw runs persistently on your own machine, Claude spins up a fresh sandbox each time and runs the skill in there.

The mechanism behind all this is Markdown. These days, Markdown is the programming language of AI. You can go more structured if you want, but Markdown is the sweet spot for squeezing the most out of an agent: you write the flow, the steps, the criteria in a plain .md file, point Claude Code at that folder, and it reads everything in. Better still, it’s universal — you can hand that same file to Codex or Google’s Antigravity and they’ll all read it. (It’s the same reason Markdown is great for your personal notes, but I digress.)

I wanted Claude to draw molecules

and Claude outdid itself

I’d never actually bothered setting up Claude skills. I knew they existed, I’d skimmed the list of available ones, but I’d never wired any up myself. That changed thanks to an exam.

I was reviewing for a test coming up in a couple of months, and it occurred to me that I really needed to know my molecule shapes. My first thought was to keep a catalog of them to refer back to. My second thought, was to ask Claude to generate them as SVGs, instead of manually downloading them myself. My third, better thought: what if Claude could just draw them for me on demand?

This is where tooling wins. You see, drawing here isn’t the same as “generating.” This is not going to be the same as an image model. It’s not going to snitch together what it thinks looks like the correct structure, and it won’t be prone to AI hallucinations.

Chemistry already has machine-readable formats for this. A molecule can be represented with a SMILES string, and libraries like RDKit can turn that into a proper 2D structure. Claude just had to learn to use those.

I’d seen Claude use dynamic, interactive widgets to explain things before (ask it about interest rates and it’ll spin up a little chart you can play with right in the chat). So I asked if it could do something similar for molecules. It could not do it out of the box, but that was not really a problem. When it told me it can’t do it, I told it “learn how to.”

When I got the first output, there were rough edges, so I told it what to fix, and it fixed them. At this point I still hadn’t even planned to make a skill — I just wanted to render molecules so I could download the SVGs and catalog them myself. Then I pushed further and asked if it could render them in 3D. It downloaded what it needed, and there they were: rotatable 3D models, live inside the chat.

Finally, I asked it to export the whole thing as a skill. It did, and gave me a tidy Save as skill button. One click, and from there the skill was usable across Claude, Cowork, and the regular app, from my browser or any device I’m signed into. With that, once I had perfected the output, I could just ask it to draw or compare structures anywhere, at anytime.

Use the expensive model to do the hard part

Once the skill works, cheaper models can reuse it

Since you’re building the skills with the LLM itself, I suggest using a stronger model to build the skill. I built mine with Claude Opus 4.8 so it’d be thorough the first time around. But, afterwards, it’s useable with any cheaper model that supports tools. When I tried to use the skill with Claude Sonnet, it failed and dropped an SVG as an artifact instead of using the inline widget. I was feeling lazy, so I asked Sonnet what went wrong, took its explanation back to Opus, said “this happened, fix it,” and Opus fixed it. Done.

Another advantage of Markdown here is that you can easily export the skill to other platforms. Codex can read it, Antigravity can read it, and OpenClaw can read it too.

Speaking of which… Claude conversations use separate sandboxes, so if a Skill needs external packages, a new conversation will need to download or set them up again. The Skill itself is saved, but the runtime is not the same as having your own machine always ready with everything installed.

That is one advantage of something like OpenClaw or Claude Code, which runs locally. If you set up the environment on your own machine, it can stay there. But Claude has the better interface for this kind of visual output. For my molecule use case, that matters more than having a persistent local environment.

This was just one use case I happened to have. Maybe I’ll set up a database, have it render all of the ones I need, so that the next times it only has to show them instead of building them. Maybe I’ll set up an MCP server. For now, I know what I’ll be spending my credit on: building more Claude skills.

But that’s the point — the sky really is the limit. You just have to ask it to learn, and then save it as a skill. Build the skill with Opus so it’s solid, format it properly, and then run it wherever you are at a fraction of the cost. The skill gives you persistent high quality output, all the while slashing costs since you don’t have to repeat yourself every time.

Developer

Anthropic PBC

Price model

Free, subscription available