By Chris Wang & Claude Code
AI can now do the execution.
The question isn't whether —
it's what humans do next.
The unit of interaction was the instruction. cd, mkdir, cp. You told the machine exactly what to do, step by step. The human held the full mental model. The machine was a precise but entirely literal executor.
"Set a timer for 10 minutes." The shift was meaningful — you no longer specified how, just what. But it was still transactional. One request, one response. No sustained collaboration, no shared context, no compound outcomes.
"Build me a meeting prep tool with live presenter mode." The machine decomposes, plans, executes across files, debugs, iterates. The unit of interaction is now the goal — and the machine has genuine autonomy in achieving it.
If AI handles execution, the human role shifts to something like a film director — providing vision, taste, judgment, and coherence. Not operating the camera. Not building the sets. But no one would say the film isn't theirs.
Direction is an active, ongoing relationship. "Warmer, more restrained, less corporate." Seeing a layout and knowing the spacing is wrong before you can articulate why. The back-and-forth is the creative process.
The barrier to directing has collapsed. A product manager who has never written code can direct creation of a sophisticated tool. Understanding the medium still makes you a better director — but the entry point is taste, not technique.
"You were never your tools. You were never your technical skills. You were always the person who knew what to make and why."
When you created a machine, you design and build a thing. When you created intelligence, you design and build the Machine that designs and builds the thing. You're no longer the architect. You're the gardener.
Traditional design is spec-driven: wireframes, requirements, pixel-perfect mockups. The new mode is simulation-driven. Define parameters — Motion, Warmth, Density — and the system explores the space of possible outputs.
You shape conditions for growth, not the growth itself. This isn't hypothetical — it's what generative design does in engineering, what procedural generation does in games. Now it's how we build software.
The terminal isn't a constraint — it's a creative medium. The file system is your canvas, the operating system your studio, the project structure your composition. Natural language is the brushstroke.
Instead of expressing intent directly, express meta-intent: the intent to build the tool that will accomplish the original intent. "Build me a motion design studio" instead of "animate this element."
Each meta-intent creates infrastructure that makes the next thing easier. The motion studio, the meeting prep tool, the design system — each one is a multiplier, not just a deliverable.
"When you created intelligence, you design and build the Machine that designs and builds the thing."
The most powerful thing you can teach AI isn't facts or procedures — it's principles. "Motion should tell a story." "Subtlety over spectacle." Teach the why, and the what follows.
General models know everything and prioritize nothing. The antidote isn't fine-tuning — it's hand-curating context to give the model taste and focus. Specialization is curation, not limitation.
Your design principles, CLAUDE.md files, session logs — they're not documentation. They're taste codified into system instructions. The AI absorbs your aesthetic and applies it consistently across every project.
Human teaches AI taste through feedback. AI teaches human domain knowledge through demonstration. Both sides get smarter with every interaction — a compounding loop that accelerates over time.
Pre-training learns the world (billions). Post-training learns aggregate preference (millions). Runtime learning learns you (one person, right now). Interactive tools are the teaching medium — every slider, every toggle encodes taste.
Session logs, design tokens, project memory — each session builds on the last. The AI doesn't just respond; it remembers, adapts, and compounds. Collaboration deepens with every conversation.
"Human teaches AI taste. AI teaches human domain knowledge. Both sides get smarter with every interaction."
The Sapir-Whorf hypothesis says language shapes thought. The same applies to interfaces. A chat box invites conversation. A node editor invites systems thinking. The interface determines what both sides can think together.
Intelligence doesn't live in the model alone. It lives in the space between human and machine — in the quality of the interface, the richness of the context, the tightness of the feedback loop. Design the interaction, design the intelligence.
Express intent quickly. See results immediately. Refine precisely. Maintain coherence. The quality of the director-AI feedback loop is the quality of the output. Every great AI product is really a great feedback loop.
Software used to be the prerequisite — build the app, then use it. Now software is the residue of interaction. You solve a problem together, and software crystallizes in the wake of that collaboration.
Voice, sketch, code, files, camera — the input space is expanding. The GUI shifts from product to channel. Interaction flows through whatever medium is most natural for the thought being expressed.
The meeting prep tool wasn't planned — it emerged from needing to prepare for meetings. The design system wasn't specified — it grew from making things look right. The best software is born from use, not from plans.
"Intelligence doesn't live in the model. It lives in the space between human and machine."
This is computational design.
Augmenting, not replacing humans.
Humans thinking at the speed of computation.
Directing, not executing. Growing, not grinding.
Time to learn and build, and be creative.