Prompting AI is hard and unnatural because the underlying protocol — the rules shaping how you interact with the system — hasn't changed since the era of punch cards. You assemble a complete request, submit it, wait for output, find what's wrong, and repeat. The AI's intelligence has exploded. The interface has not. And when the magic words don't land, people blame themselves. They shouldn't. The mismatch is the interface, not the user.
Why Does Prompting AI Feel So Hard and Unnatural?
Think about the last time you typed a request into an AI chatbox. You probably rephrased it at least once. Maybe you added more context, tried a different angle, or searched for a better way to word it. That friction has a name, and it isn't user error. It's a design problem baked into the oldest protocol in computing history.
02:15
Ted Johnson introduces the core disappointment behind founding Join an AI — why powerful AI still feels unnatural to use
Watch at 02:15 →
Ted Johnson, co-founder of Join an AI, has spent 25 years building enterprise software and human-centered interfaces. When ChatGPT arrived, he felt two things simultaneously: amazement that the world had changed forever, and a surprising, stubborn disappointment he couldn't shake. That disappointment became a question that launched a company: why do we still have to learn AI? Why does something this powerful feel so unnatural to use?
The answer, it turns out, isn't about model capability. It's about what happens between the human and the model — the interface itself.
Is Prompt Engineering a Real Skill or a Workaround?
Prompt engineering is often sold as a superpower — the insider knowledge that separates people who get great results from AI and those who don't. Tell it to think step-by-step. Give it examples. Ask it to act as an expert. Don't ask it that way. Paste more context. Use markdown. We trade incantations like they're wisdom.
But strip the flattering label off and prompt engineering is just a set of rules for packaging up a batch job. It's the same mastery a punch card operator had — knowing exactly how to assemble the deck so the job wouldn't fail. That's not a power user skill. That's a workaround for a protocol that hasn't kept up with the intelligence it's supposed to serve.
The uncomfortable truth: we've gotten good at prompting these black boxes, and that's the part that should bother us, not reassure us. Punch cards weren't bad. Command lines weren't bad. They were brilliant solutions to the constraints of their time. The question is whether batch is still the right protocol — whether we're still pre-packaging our intent for a machine that no longer needs us to.
Channel, Expression, Protocol: The 3 Keys to AI Interfaces
To understand why the prompt is still a punch card, you need three concepts that explain how any interface actually works.
08:40
The punch card protocol explained — batch submission, wait, read, fix, resubmit — and how the modern prompt mirrors it exactly
Watch at 08:40 →
The Channel
The channel is the physical medium the interface gives you to work in. A keyboard is a channel. A microphone is a channel. A prompt box is a channel. Each one can physically carry a different kind of signal — text carries discrete symbols, voice carries timing and pitch and hesitation, a diagram carries spatial relationships all at once. Channels matter because they define the raw bandwidth of what can be transmitted.
Expression
Expression is how much of what you actually mean will fit through the interface. Early computing was brutally limited here. Assembly language gave you a few dozen opcodes. Shell commands gave you flags. Programming languages gave you composable primitives. Each was a fixed vocabulary — you expressed your intent by choosing from a menu the machine would accept.
Natural language blew that menu open entirely. For the first time, you can say almost anything the way you'd say it to another person, and a machine can take it in. On the expression axis, the leap AI delivered is real and enormous. There's an ocean of meaning in an ordinary human request — context, nuance, intent — things we've never had to spell out to each other. Now we don't have to spell them out to machines either.
The Protocol
The protocol is the shape and rules of the interaction itself. And this is the piece that hasn't kept up. The channel stayed the same for 180 years. Expression exploded in the last three years. But the protocol — prompting — is still the protocol of a punch card.
Why Are LLMs Still Stuck in Batch Mode?
Punch card batch had a very specific shape: you sat away from the machine, carefully encoded your entire request in advance, submitted the job, waited — sometimes overnight — read the printout, found one thing wrong, fixed it, and resubmitted. The machine never engaged with you while you were thinking. It engaged with the finished package after the fact.
Now look at the prompt. Assemble the whole request. Submit it. Wait. Read what comes back. Something's off — assemble it again, submit again, wait again. The wait shrank from overnight to a few seconds, and that speed fooled us into thinking the interaction had become truly conversational. It hasn't. It's still batch. You still package a complete turn before the machine is allowed to participate.
14:22
Nvidia Personal Plex demo shows real-time interruption handling — the AI stops mid-sentence, yields the floor, and resumes naturally
Watch at 14:22 →
And here's the historical irony: batch computing inherited its protocol from the weaving loom. You set the entire pattern in advance, then ran the cloth. Punch cards borrowed it. Computers inherited it by default. And now AI — the most capable reasoning system ever built — got handed the protocol of a loom.
The model can ask a follow-up. It can clarify mid-thought. It can notice it's missing something and say so. It should be doing all of these things. Instead, we're still making people submit the deck and wait for the run.
16:50
Live demo of a group meeting where an AI follows conversation context, identifies speakers, and contributes without being explicitly prompted
Watch at 16:50 →
How Does AI Interface Design Need to Change?
Model capability is shooting straight up — reasoning, speech, vision, memory, planning, all improving fast. The interface protocol is flat. And when the gap between those two curves widens, the human absorbs all the friction: deciding what context matters, remembering what to ask, choosing the timing, noticing the ambiguity, repairing the output, engineering the prompt.
The intelligence feels magical. The interface still feels like work. And when it feels like work, when the output is wrong, people blame themselves. They decide they're bad at AI. They're not specific enough. They just don't get it.
It is not the user's fault. We are being asked to operate a brand new kind of intelligence through the protocol of a punch card. The mismatch isn't the user — it's the interface.
The right design question isn't "how do we make prompting easier?" It's: what burden are we still putting on humans only because the machine used to be too limited to carry it? Ask that question, and the whole interface space opens up.
Can AI Join a Real Conversation Without Being Prompted?
The answer is yes — and it's already starting to happen. OpenAI's voice mode now backchannels, producing the small listening sounds ("mm-hmm," "right") that signal active engagement in human conversation. Nvidia's Personal Plex research model handles real interruptions — it stops mid-sentence when a person cuts in, yields the floor, and picks the thread back up when the moment is right. That's genuine turn-taking: listening and speaking at once, in real time.
But the deeper challenge isn't just conversational flow. It's understanding the social context of a conversation — who's speaking, whether words were meant for the AI or for someone else in the room, when to take a turn and when to stay quiet. A voice assistant that hears "Hey, come on in" directed at a colleague in the doorway and responds "Sure, what's on your mind?" isn't dumb. It's working perfectly within its protocol. The protocol just doesn't have a slot for that kind of human context.
The more ambitious vision is an AI that sits inside a group conversation — a meeting, a working session, a design review — following along, understanding who holds the floor, labeling what's being said (question, proposal, answer, aside), and choosing its moment to contribute without being explicitly summoned. No one writes a prompt. No one hits submit. The system is in the conversation, and it participates the way a thoughtful colleague would.
What Does Human-Centered AI Design Actually Look Like?
It doesn't mean making everything voice. It doesn't mean replacing human judgment. It doesn't mean a wall of markdown or a decade-old set of digital constructs. The right answer is the affordances humans already use with each other: a question, a pause, a sketch, a checklist, a quiet aside — or saying nothing at all.
For 75 years, humans adapted to the machine — its syntax, its forms, its timing, its batch protocol. A system that can reason, listen, infer, and adapt should be able to meet us partway, if not all the way. AI is not just an intelligence technology. It's increasingly an interface technology. And book-smart models alone are not enough.
Computing has mostly been about improving how humans encode intent for machines. The punch card, the command line, the mouse click, the iPhone swipe, the prompt — every step was real progress, and every step carried the old constraint forward into the next era. A translation tax. A precision tax. A context tax. A repair tax.
AI is the chance to put those down. Not by making everything magical, but by making computers, for once, more fluent with us. Because if a machine can finally understand more of what we mean, then we can — and should — stop reshaping ourselves to be understood by it.
