Domain Skills Are Coming. Start Thinking About Them Now.

Speculation: This post is based on observations from the Claude Chrome extension's source code and a direct conversation with Claude in Chrome. Anthropic hasn't announced any of this as a product feature.

A few weeks ago I asked Claude in Chrome a fairly direct question: do you have access to any domain-specific guidance based on where I'm browsing? The answer surprised me.

"Yes. In the tab context I receive, there's a domainSkills field that can contain guidance and best practices for specific websites."

That's a direct quote. And while it sounds like a small technical detail, I think it points at something that has been missing from the AI-assisted design workflow conversation for a while.

What domainSkills actually is

A source code analysis of the Claude Chrome extension (v1.0.56) gives us a clear picture of the architecture. When you send a message from the extension's side panel, Claude doesn't just get your question and the current tab - it also receives domain-specific guidance pulled from Anthropic's servers. The extension queries an endpoint at api.anthropic.com that classifies domains and, when available, returns structured guidance for that specific site.

This is how the extension already knows how to navigate Gmail or interact sensibly with GitHub. That knowledge isn't just baked into the base model - it's delivered contextually, via domainSkills, based on where you are.

Right now, this is entirely controlled by Anthropic. You can't create your own. The content is curated internally and covers popular consumer platforms. If you're on mail.google.com, you get Gmail-specific guidance. If you're on github.com, you get GitHub-aware context.

That's genuinely useful. But it's also pointing at a much more interesting question.

The gap that's been bothering me

There's still no good answer to a problem I hear from design teams fairly regularly: how do you teach an AI to work meaningfully within your specific product?

Not generic UX guidance. Not broad design principles. Your navigation model. Your component library. Your error handling patterns. The things that a new designer on the team needs three months to internalise - and that an AI assistant currently has to pretend it can infer from screenshots or documentation you paste in by hand.

Some teams try to solve this with long system prompts in their internal tools. Others build custom GPTs or Claude Projects with uploaded docs. These approaches work, up to a point. But they're brittle, don't scale across contexts, and require someone to actively maintain them. The AI doesn't know your product - it knows whatever you remembered to include in that session.

Domain skills, if opened up to user or team configuration, could be something fundamentally different.

Why this architecture matters

The key thing about how domainSkills works in the Chrome extension isn't the content - it's the delivery mechanism. Guidance is attached to context automatically, based on the domain you're on, before the conversation even starts. You don't have to prompt for it. You don't have to paste anything in. The AI just knows what it's dealing with.

Now imagine that same mechanism, but configurable - in two distinct directions.

The first is internal. You associate guidance with a URL pattern - *.yourproduct.com, say - and that guidance gets injected into any AI-assisted session on that domain. Your component library conventions, your IA patterns, your accessibility requirements, your error state vocabulary. All of it available automatically, in context, whenever a designer or developer is working in or adjacent to your product. The AI doesn't need onboarding. It already knows your conventions before the first message is sent.

The second direction is more interesting, and I haven't seen anyone talk about it yet. What if you could publish domain skill guidance for your own product - not just for your internal team, but for anyone using Claude on your platform? If you're building a SaaS product and your customers are increasingly using AI assistants alongside your app, you could author the guidance that shapes how Claude understands your interface. Your navigation model, your terminology, your key workflows - hosted by you, served automatically when anyone with the extension lands on your domain. You'd be giving Claude a manual for your product, written by you, available to every user without them having to configure anything.

That's a genuinely new distribution model for product documentation. And it reframes something that's always been a support cost - teaching users how to use your product - as a design artifact with real leverage.

What designers should be thinking about now

This is speculation, to be clear. Anthropic hasn't announced anything. There's no indication of a timeline. The feature might open up, or it might stay locked to Anthropic's own curation indefinitely. I genuinely don't know.

But if it does open up - even partially, even for enterprise tiers - teams that have already done the thinking will have an advantage. And doing the thinking isn't as complicated as it sounds.

The useful question isn't "what would I put in a domain skill?" It's "what does a new designer on my team need to know by week four that they couldn't have guessed?" That's your starting point. The patterns that aren't obvious. The conventions that exist for a reason but aren't documented anywhere except someone's head. The tradeoffs you've already made that shouldn't be re-litigated every time someone touches a particular part of the product.

For teams building customer-facing products, the question is slightly different: "what does a user with Claude open need to know that our onboarding doesn't cover?" That's also worth writing down. Even if domain skills never materialise the way I'm imagining, you've still done useful work.

That knowledge exists in every mature design team and every mature product. The problem has always been capturing it in a form that's useful rather than just stored. Domain skills could be the format that finally makes that worth doing.