If you think the tools and platforms you depend on have gotten too complicated to keep up with, you're right. And AI is starting to look like the way through.

That's the whole argument. The rest of this is why I think it's true, and where I think the answer is taking shape.

The platforms you're already operating

A short list of the platforms a fairly ordinary independent professional now has to operate competently:

Each of these is, individually, a deep domain. Substack alone has dozens of settings most users never touch — embed mechanics, email throttling, pledged subscriptions, sections, restacks, cross-posting, audience segments, scheduled drops. Cloudflare has hundreds of products. The AI coding agents have competing instruction-file standards that change every six weeks.

You aren't supposed to be an expert in all of these. You weren't trained for it. You picked up a few because the work demanded it, and you're getting by on the rest with shallow operating knowledge and a healthy fear of clicking the wrong button. The pieces that "just work" are doing so because someone you don't know patched a bug last Tuesday.

This is a real load. It is not a personal failing.

The cost has a name

Call this the expertise tax. Every platform you use individually charges you in study, experimentation, and small mistakes — most of which you absorb without noticing because they're paid in five-minute increments instead of in dollars.

The expertise tax is foundational. There is also an integration tax — the cost of working across platforms when none of them know about each other. You publish a piece. It needs to land on your site, then on Substack, then on LinkedIn three days later, with email on the flagship and not on the cluster, and the timing is yours to handle because no platform will do it for you. That tax is real and large. But it sits on top of the expertise tax, not next to it. You can't coordinate platforms you don't know how to operate.

What's changed in the last two years isn't that the tax went up. It's that we now have something that can pay it on the user's behalf.

Why this is new

Manuals, tutorials, courses, and screencasts have existed forever. They transmit expertise — but they require the user to acquire it before acting. You read the manual, then you do the thing.

That model has been the only model. Until now.

What's actually new is that a sufficiently capable AI can hold encoded expertise in its context AND execute on it without the user ever acquiring it themselves. The expertise sits in a file the model reads. The action sits in a tool the model calls. The user speaks intent in their own language. They don't need to know the platform's vocabulary, settings, or quirks. They need to know what they want.

That's a category of capability that didn't exist five years ago. It's not a better manual. It's a manual that reads itself and acts on the reader's behalf.

This is the load-bearing claim. If you don't believe AI can do this reliably yet, the rest of the argument is interesting but not urgent. If you believe it can — even partially, even in narrow domains — then the question becomes what shape the answer takes.

The four-aspect shape

I've been working on the answer's shape across several products and articles for a while. The clearest cut I've found is that the layer that pays the expertise tax does four things, and most useful products do at least two of them at once.

Knowledge. Encoding what platforms exist, how they work, where the gotchas are, what changed last week. The CLAUDE.md and AGENTS.md files I wrote about earlier this week are knowledge artifacts — they tell an AI tool how to operate inside a specific tech stack without the user re-explaining every session. The kits and starter packs proliferating on GitHub right now are the same shape.

Interpretation. Translating the user's intent — said in plain language — into the platform's native action language. "Schedule this for next Friday" becomes the right buttons clicked in the right sequence. "Make my site easier for AI to read" becomes a set of structural changes a non-technical user couldn't articulate but would recognize as right when they saw them.

Coordination. Executing actions across one or more platforms in the right order, at the right time, respecting cooldowns and dependencies. Single-platform coordination handles the within-platform sequencing nobody wants to do by hand. Multi-platform coordination is what most people mean when they imagine AI agents — but the single-platform version is more common and almost as valuable.

Memory. Holding state, doctrine, history, and intent across sessions, surfaces, and time. What the user has done. What the user has decided. What the user prefers. Without memory, every session restarts the relationship from scratch and the user pays the tax again.

Most products you'll see in this category are blends. Coordination plus Memory, or Knowledge plus Interpretation. The pure-aspect cases are rare. The four-aspect cases — the products that do all four well — don't really exist yet at consumer scale, but I think they will, and within a few years.

A few of us have started calling this the Third Layer

The name is provisional. It might not stick. But the category is real whether the name lasts or not.

The framing is simple. The platforms are layer one. The AI itself — the model, the assistant, the tool — is layer two. The Third Layer is what holds the user's intent across both. It encodes the expertise the platforms demand. It lets the user operate in their own language instead of in the platforms'.

If you want to test whether you've felt this category before: think about the moments you've used an AI assistant and wished it knew the things you already know about your own setup — your preferences, your projects, your platforms, your prior decisions. That gap, between what the AI can do in general and what the AI can do for you specifically, is where the Third Layer lives.

Where this is going

We are early. The products that do this well are mostly small, mostly DIY, mostly invisible to people who aren't already deep in AI tooling. The kits, the templates, the starter packs, the brain documents, the pacing protocols, the cross-platform orchestrators — all of them are first drafts of a category that will eventually have polished consumer products in it.

The interesting question isn't whether this layer exists. It exists. The interesting question is what happens when it gets good. When the encoded expertise stops decaying because the layer maintains itself. When the AI starts noticing patterns the user hadn't articulated. When the layer doesn't just translate user intent into platform action but starts shaping what the user thinks to ask for in the first place.

That's a longer argument for another piece. For now, the smaller claim:

If you've been feeling the load and assuming it was your problem to solve through harder work, I'd like to register the alternative view. The problem is real. The cost is significant. And the answer is closer than it looks.

The tools and platforms got too complicated. AI is starting to look like the way through.