The Machine is Ready. Now Write an Essay About Yourself.

You open ChatGPT, the cursor blinks and you genuinely don't know what to type. This happens to everyone, including the people who publish prompting frameworks about it. The standard explanation says you haven't learned to prompt, that would be comforting, because that can be taught.

The model's ignorance of you can't be phrased away. It doesn't know whether you run a bank or a bakery and a perfectly worded question sent to a system that knows neither returns the same advice everyone else gets.

The industry knows this and it has been shipping fixes for two years.

The Patches

The first fix was prompt suggestions, those little buttons under the text input field proposing what you might ask next.

A menu beats a blank page. But it's the same menu for every human on earth and it is weak exactly where AI would be most valuable, because it is generic and safe. No suggestion button has ever read „Help me prep for the conversation with my underperforming account manager.“

The second fix was memory. This has been added to about every major AI app and it actually learns things about you. Unfortunately it hoards a lot of things with no sense of what became irrelevant months ago. It will sometimes confidently suggest next steps for a project that died in March. Personal-and-stale is even worse than generic, because a wrong assumption applied with confidence is just a waste of time and tokens. (And sure, you can edit memories, assuming AI database administration is your newfound passion).

Put the two fixes side by side and the lesson writes itself: One knows the questions but not the person. The other knows the person but not the present. Context isn't just something you store, but something you select in the moment.

The Knowledge That Just Became Usable

There's a reason this is a topic now rather than ten years ago, because every organization runs on tacit knowledge nobody writes down.

An experienced sales director knows intuitively who actually holds the power in a room, not necessarily by their job title, but because everyone else subtly glances at them before answering a hard question. For fifty years, this knowledge was completely worthless to software. Legacy software digested numbers and structured fields. We were forced to compress human reality into rigid dropdown menus and the interesting, nuanced parts of the meeting didn't survive.

Documentation became a tax paid by the one person who doesn't need the record (the employee), coming due at the exact moment the knowledge feels too obvious to write down. As a result, the most expensive database in any B2B company is the CRM and it's full of fiction.

But language models run on language (really). Human reality is is an unstructured mess and for the first time, computers can work with that. Getting it out of people's heads, however, requires a completely new approach to interface design.

Proof Of Concept

Here's the situation where I kept running into this problem: an IT vendor tells a room of executives that they can build any AI application imaginable. An executive answers, sincerely, that this is wonderful, but they don't know what their use cases are. Both sides are stuck.

The use cases probably exist in the executive's tacit knowledge of their own work. But if they won't articulate them to another person, they definitely aren't going to type them out to a machine. When they go back to their desk, the blinking cursor demands they write an essay about their workflow just to get started.

At the moment, AI relies on these thick interfaces that demand high cognitive effort. In UX psychology, asking a user to generate an answer from a blank slate is called recall and it is exhausting. So, I built a proof of concept of a prompt generator designed around recognition, which costs the brain almost nothing.

It's called The AI Deck.

You swipe through twelve cards on your phone, each one a statement about how you work, agreeing or disagreeing as you go. The whole experience takes about ninety seconds. The deck never asks you to produce context about yourself; it only asks you to recognize it. Behind every card sits a prompt snippet I wrote months in advance. Your swipe silently adds that sentence into the final prompt. You can then copy and paste it into any AI app and it generates use cases tailored exactly to how you work.

Compare the two experiences. Without the deck, our executive types "What can AI do for me?" and receives a generic ten-point list. With the deck, the AI opens by talking about them: noting they watch carefully before budgeting risk, their expertise in SME lending and their preference for voice dictation.

One card is called "Build instead of presenting." If they swipe yes, the AI knows they would prefer to skip slides and generate simple dashboards for future research presentations. Five minutes later, they are discussing a small dashboard for a quarterly review with ChatGPT. They never typed a word about quarterly reviews or slides. They recognized one sentence within the card and the recognition did the explaining for the AI.

The Toolbox Opens

The AI Deck taught me something subtler than "swipes work." Twelve binary cards are great for catching the static baseline of a person, but daily work isn't static.

When our sales executive walks out of a messy negotiation, they don't need a swipe deck. They know exactly what just happened: the CRM says the VP is the decision-maker, but during the pitch, everyone kept glancing at the quiet IT director before answering hard questions. The lowest friction tool to capture that unsaid power dynamic is probably a 30-second voice dictation inside a taxi.

The trick is to combine the static (the swipes) with the live (the voice).

The AI model acts as a dynamic intelligence, reading the taxi rant through the structural logic we built during onboarding. It cross-references the messy voice note with the rigid numbers in the CRM. And then, it does something legacy software could never do: it dynamically generates the next interaction based on the exact depth the user needs.

If a piece of specific data is missing, it pings the exec with a targeted multiple-choice swipe question generated specifically for that deal. If they want to strategize, the chat asks: „I updated the CRM to flag the IT director as the true blocker. Do you want to brainstorm a game-theory approach on how this might play out in our favor?“

The interface shapes itself to the user. We have finally replaced the blinking cursor with an engine that offers infinite complexity if you want strategy and invisible simplicity if you just want to dump data.

The AI Design Challenge

Building that engine is why the blank box problem is misdiagnosed as a pure IT project. Wiring the CRM, the transcript and the language model together securely is a serious engineering job. The missing half is deciding what human knowledge is worth extracting in the first place.

If you build software, this is now your challenge too. Your users hold incredibly valuable tacit knowledge, your AI agent desperately needs it and your interface likely greets them with the exact same blank box this article opened on. Deciding which context matters and choosing how to extract it without friction, that is the work.

The AI Deck is my own playground for this. It takes two minutes and you'll get a prompt that helps you start brainstorming with the AI app of your choice.

Give it a spin: https://aideck.huszarconsulting.net/

If it leads to a good idea, or even if you absolutely hate it, I'd love to hear about it. And if you're wondering what this would look like for your product, I'm easy to find.

By Daniel Huszár