What I Actually Think About AI Platforms

Every platform converged on the same pattern: text in, text out, history on the left. Users are conditioned to it now. That conditioning is a product constraint. Whatever any of these platforms build above the chat line has to fight the gravitational pull of the interface people already know.

The result is a two-tier problem none of them have solved. Regular users stay at the chat window and get something useful. Power users climb into Projects, agents, memory systems, and code tools and get something genuinely powerful. The distance between those two tiers is growing. The platforms are building faster than they are teaching, and the crossing is invisible.

I am going to call this the Capability Gap. It is the design problem underneath all the benchmark noise.

Claude is the platform I use most heavily, and I want to be precise about why. It is not because it is the best platform overall. It is because it is the best platform for building and managing structured work across sessions. Projects, Memory, Claude Code, Skills: when that ecosystem is configured correctly, it is genuinely powerful in a way the others are not. The context window is large. The output on long, technical work is consistent.

The word "artifact" is a tell. Eighty percent of Anthropic's revenue comes from enterprise and developer workloads, and thirty-six percent of Claude usage is coding tasks. Claude was built for technical users first and that choice is visible everywhere in the interface language. That is not a criticism. It is context.

The UX is clean and deliberate. Anthropic made a considered decision to keep the interface quiet and let the content lead. I think that is right.

But there is a real failure at the context layer. I have been using Claude for over a year. I found the indicator that tells me I am inside a project chat tonight. There is a small icon. It was not doing enough work for me to notice it for a year. To be fair, it may not have existed until recently. But that is almost the point: when a senior designer who uses the tool daily cannot reliably orient herself, that is a product design problem, not a learning curve problem. You should always know where you are. What the system remembers. Where the context resets. That signal needs to be persistent and impossible to miss, not something you find by accident.

Token limits shape behavior in ways the platform probably does not intend. Opus forces heavy usage fast and the ceiling arrives quickly for anyone doing serious work.

ChatGPT earns something the others do not: it speaks plainly. It says "photo" where Claude says "artifact." That is a product philosophy expressed in a single word choice and it matters enormously at scale. When you are building for users who are not technical and never will be, plain language is not dumbing down. It is respect.

The platform is the most flexible in terms of connectors and reach. It is also the default verb for most people who have heard of AI but do not think about it much. That habit is a moat.

The ceiling is less interesting for complex workflow design. The power features require the same kind of intentional climb that Claude's ecosystem does. The floor is lower. The gap is still there.

I built my first production apps using Gemini because at the time, for design and HTML work, it gave me a better first pass than anything else. That is still true for certain tasks. Ask it for a strong starting point and it delivers.

The problem is what happens next. Additional passes start falling apart. I have tested this recently and the pattern holds. It is not that Gemini gets worse. It is that the degradation is unpredictable, which makes it hard to rely on for sustained work.

The platform is the most visual of the three and the most ecosystem-connected. NotebookLM is the standout piece: source transparency built into the core experience, not added as a feature. When I need research grounded in actual documents with visible sourcing, that is where I go.

The side navigation is where Gemini loses people. Once a user moves past standard chat into Gems or AI Studio, orientation breaks down fast. Feature-rich and easy to navigate are not the same thing. Right now Gemini is one without being the other.

Perplexity is not competing on the same axis as the others. It is a research tool and it is the best one. Every answer is visibly grounded in sources you can read yourself. That is a product decision that builds a different kind of trust.

For enterprise contexts where auditability matters, where a user needs to know not just what the answer is but where it came from, that philosophy is directly relevant. The other platforms are getting better at citations. None of them have made it the center of the experience the way Perplexity has.

Here is what I actually observe when I think about non-power users in enterprise settings. They are not giving up and walking away. They are opening the chat window, typing a question, getting an answer, and closing it. Every day. And meanwhile there are tools in that same platform that could meaningfully change how they work, and those tools sit untouched because nobody made the path to them legible.

There is also a trust problem that none of the platforms are addressing directly. I was hesitant to set up Claude's Cowork and give it access to my files. I was hesitant to give Claude Code the same. I am a designer who thinks about AI systems professionally and I still paused. That hesitation is real, it is rational, and it is not unique to me. The platforms are asking users to extend trust to systems they do not fully understand yet, and the onboarding does not do enough work to earn that trust incrementally.

Not ten ideas. Two.

State visibility first. Every user on every session should know exactly where they are. What the system remembers. Whether they are in a project or a fresh context. What will carry forward and what will not. This should be a persistent, readable signal: not an icon you find after a year of daily use. State ambiguity is a trust problem before it is a UX problem. When users are not sure what the system knows, they compensate by over-explaining, repeating themselves, or avoiding the feature entirely. Fixing visibility fixes behavior downstream.

A capability path, not a feature list. The platforms release features and wait for users to find them. The better model is surfacing the next useful thing based on what someone is actually doing. If a user has been in chat for thirty days doing structured work, show them what Projects would add to that workflow. Not as a tooltip. As a moment in the actual experience. The goal is not to push features. It is to close the distance between what a user is doing and what they could be doing, in a way that feels like a natural next step rather than a product announcement.

These are not novel ideas. They are interaction design fundamentals applied to a layer where most platforms are still thinking like engineers. That gap is where the design work lives.