On using AI in product development

As a child I dreamed of computers that could talk. One of my first programs was a naïve CLI chat written in Pascal. It might have looked like this:

program Computer;
var s: string;
begin
  repeat
    readln(s);
    if Pos('hello', LowerCase(s)) > 0 then writeln('Hello, Rodion!')
    else if Pos('how are you', LowerCase(s)) > 0 then writeln('All good!')
    ...
    else writeln('I don''t understand :(');
  until False;
end.

When ChatGPT launched in 2022, for me—someone who devoted his whole life to technology—it was pure excitement. I can finally talk to a computer! Humanity created a new kind of intelligence! (Or did it always exist and we just discovered it 🗿, the way π existed before we found out?)

Progress is crazy fast. Every few weeks a smarter model is released, or the same intelligence gets cheaper. The bottleneck is no longer the models themselves—it is the ability to harness them and build features people actually want. Apart from the new, truly AI-native companies (say, Cursor), most of us, including Poster (my company), have not yet used the potential of models that were SOTA last autumn.

Building is almost too easy now. Every other person I know is vibe-coding something. And yet where are all those new high-quality products? Is the world finally free of 🫄🏻 Electron bloatware? No. What I see is a lot of junk software and awful social media content. Why?

Watching every season of Drive to Survive does not make anyone an F1 driver. While daily driving Fable might give a feeling of being a Verstappen, it is experience that lets you steer the model to create great software. Without guidance, LLM will produce a mediocre result.

To make great products you need to develop an intuition to make decisions across different domains: programming, UX, unit economics, distribution. That intuition is what the model multiplies. So, experienced builders get the best leverage ever, and everyone else gets, well, a magic box that creates something.

But this comes with a cost. LLMs hand you superpowers, but sometimes you need to actively defend against them to stay in control.

Don’t delegate thinking to AI

LLMs can create the median of what has already been done before. You can create something new and truly creative.

I like to use a “reversed data flow” when talking to AI. Instead of asking questions, I would often tell it to ask questions of me—and I find gold in my own thinking about the answers: ideas only I could stumble upon, with my experience and context.

A five-word prompt, “ask me clarifying questions relentlessly” (simplified version of /grill-me skill), significantly improves code quality and reduces time I spend arguing with the model.

Read the code

People who say we no longer should read the code a model writes either don’t understand how production codebases evolve, or are financially motivated for you to burn more tokens.

Sure, the way I look at code has changed. It is now more about properly articulated code style that the model should pick up (I like functional programming), right separation of concerns, architectural oversight.

But you cannot skip this. Prior to LLMs, having a mental model of the codebase was a byproduct of typing the code symbol by symbol with your own fingers. Now, building and maintaining that mental model takes deliberate effort.

Throw away what you’ve built

My relationship with code has changed—I have no attachment to it and no feeling of sunk cost. Since code is written fast, I don’t mind throwing it away. I can spend a few hours prototyping a new feature just to tinker around with it.

And then I would revert changes without regret, just because I didn’t like it, or the timing is not right yet. It lets me glance into the product’s future, and come back safely.

Don’t rush

Development speed is going up, and there is a great temptation to ship more features. But if the constraint on a product’s growth were the number of features, why did Evernote and Winamp die?

As release velocity rises, so does the rate at which mistakes accumulate. Experiments with “software factories” run entirely by AI agents turn into a dump within 3–6 months. It used to take 3–6 years to get your monolith into that state. I know, I’ve been there.

Are we raising broiler chickens or growing an orchard? In the first case raw speed wins.

Talk to people

“You’re absolutely right!” from Claude was last year’s meme. AI sycophancy has been described many times over—like other consumer platforms, AI chats are optimised for engagement and time spent. And what better way to keep you around than to convince you your ideas are brilliant?

In reality your ideas are probably bad, most of them. To find out whether the latest one will become a million-dollar business, it is better to test it with a potential customer than with a chatbot.

I have 20 chats with most engaged customers in Telegram where I drop questions every few weeks for validation.

Keep learning

Maybe my company’s products will be unnecessary in five years. Maybe product development will be at last solved. But I am certain an educated person will be able to adapt and thrive, regardless of what AI progress brings.

My learning list:

• Read books
• Learn languages (ChatGPT Voice Mode has been an unusually good unlock lately!)
• Write. Writing is thinking.
• Develop soft skills: negotiation, public speaking, try smiling more 🙂
• Learn to ask the right questions, read “The Mom Test”
• Learn humanities: history, philosophy, psychology

Rest

The feedback loop has become dramatically faster. You are able to create so much more. One prompt—a new product or feature. I know. You’re just like Bradley Cooper in “Limitless”.

Feedback that fast is intoxicating: burnout accumulates far more quickly, and spending an evening with agents paused feels like wasted time.

Get some rest, spend time in nature. They say one fellow discovered some important laws while resting in an apple orchard 🍏