An interactive essay by Chris Wang & Claude Code
We've been talking to AI wrong.
The answer is more than just better words.
You type a paragraph of what you want. The model guesses what you mean. Signal-to-noise is catastrophic — like whispering instructions across a crowded room.
Chain of thought. Few-shot examples. System prompts. Better, but you're still describing what you want instead of showing it. The ceiling is language itself.
CLAUDE.md files. Memory banks. Retrieval-augmented generation. The model knows more — but memory without agency is just a briefing. Knowing what to do and being able to do it are different things.
The model can act — run tests, read errors, iterate. Boris Cherny's insight: "2-3x quality improvement" when the AI can verify its own work. The loop closes. Intent becomes action.
David Luan's "Gym" — give the AI the entire desktop. Every application, every pixel, every interaction. The environment is the prompt. No translation layer. No information loss.
The best prompts aren't sentences.
They're systems.
A developer building tools for developers. A researcher building autonomous agents. Different starting points. Different domains. Same conclusion.
Not prompts — codified workflows. /commit, /review-pr, /deploy. Each command encapsulates domain knowledge that would take paragraphs to describe. The vocabulary grows as the tool matures.
Each agent owns a domain and a toolset. The code reviewer knows about linting. The test runner knows about frameworks. Specialization beats generalization — just like in human teams.
10-15 Claude Code sessions running simultaneously. Cognition scheduled like compute. The bottleneck isn't the model — it's how many tasks you can feed it at once.
The AI runs the tests. Sees the failures. Reads the stack traces. Corrects itself. 2-3x quality improvement — not from better prompts, but from the ability to check its own work.
The first generation: narrow, brittle API wrappers. Each integration hand-crafted. Coverage: a thin sliver of what the computer can do. "Zero tasks have pure APIs" — most real work touches the messy visual world.
The model can see the screen. OCR, element detection, spatial reasoning. Suddenly every application becomes accessible — not through its API, but through its pixels.
Record what humans do. Ghost cursor traces. Click patterns. The AI learns not from instructions but from observation — the same way humans learn most complex tasks.
The boundary dissolves. The AI has the same interface a human does — the entire screen. The environment is the prompt. No translation. No information loss. The "gym" where agents train on any task.
You don't prompt an orchestra.
You give each musician an instrument and a score.
The bottleneck isn't the model. It's the interaction layer. And that layer is mostly text boxes.
The opportunity isn't to write better prompts. It's to build better infrastructure — the scaffolding that lets intelligence act on the world.
The left side waits. The right side acts.
Goals feed tools. Tools produce outputs. Verification catches errors. Iteration refines the result. The cycle accelerates with each pass — not because the model gets smarter, but because the feedback gets tighter.
The model is the engine. But the vehicle is the entire ecosystem of tools, verifiers, context providers, and feedback loops that surround it. The quality of the output depends on the richness of this layer.
Better prompts won't get us there.
Better models alone won't either.
The breakthrough is in the space between.
Give it tools, not words.
Build the thing that builds the thing.