Eugene Tsarkov (@etsarkov) header
Eugene Tsarkov (@etsarkov) profile picture

Co-Founder & CPO building AI-native products from 0→1 to scale. Writing about product strategy, monetization, GTM, leadership and mentoring.

Serbia
Joined February 2010
Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)Photo by Eugene Tsarkov (@etsarkov)
You ask AI for help with a task, get a generic answer, and end up explaining what you meant. I built SetupKit, a free guide for situations like this: setupkit.ai/ It's for people who are new to AI or mostly use it to ask questions and get answers, and are still figuring out how to make it useful in their work or everyday life. Say you have notes from a meeting and need an action plan. You could ask AI to "make a plan," or give it a little more to work with: "Turn these notes into a draft action plan for the team. Keep it brief. If it's unclear who's responsible for something, ask me. I'll review the plan before sending it." Now the assistant has a clearer idea of what you need, who it's for, and what to ask about. SetupKit helps you work through that conversation with your own AI. It guides your assistant to ask about your goals, how you like to work, and which actions need your approval. You can then put those agreements to use on a real task. You can keep a short working brief and bring it into your next chat, so you have the context ready to share. If you're coming back after a break, a summary can help you pick up where you left off and decide what to do next. If you already have instructions and notes, you can start by reviewing what's useful, where the rules conflict, and what needs organizing. Files and folders are optional; you can begin with something useful in the chat itself. To get started, copy the start prompt from the site and paste it into your AI chat. You don't need to learn how to write prompts first. You'll still need to check the assistant's answers. Start with one small task where you can judge the result yourself. --- If you have ideas for improving SetupKit or would like to share your experience using it, I'd love to hear from you. If you know someone who could use this, feel free to send it their way. SetupKit is free, and I'd be glad if it helps them feel more confident using AI.
2
105
Asked "what are you working on right now?", someone whose product is under NDA or still in stealth mode usually has two answers, and neither is very comfortable for them or for the person asking. They can say "I can't tell you," and then the other person has nothing to rely on. Or, to sound convincing, they name the niche, what makes the product different and when it launched, all at once, and from these details people can most likely already recognize the product. But most often the one asking wants to understand the person more than the product: how they think and what they are responsible for, and for that the name is hardly needed. Without the name, the person can still share three things: 1. What problem the product solves and for whom. Companies or people, what kind of audience, with no name and no details of how it works. 2. What stage it is at. First users, first money, or already repeat sales. 3. What they decided themselves. Which decision was theirs and what changed after it. Numbers can be rounded or added up across several projects: "a few thousand active users" or "a few dozen paying companies." Before the conversation, it seems worth checking not each detail but pairs of details: the audience alone reveals nothing, and neither does what makes the product different, but together they can point to one product. Then what must stay private stays private, and the other person has enough to form their own opinion. #EugeneOnProduct
1
28
When a company looks for a product leader for a product where AI does a noticeable part of the work, the conversation almost always has the right words in it: agents, models, "AI-first". Almost everyone knows them now, and they no longer make it easy to tell someone who built things from someone who was nearby. A different question seems to say much more: where did you decide not to use AI, and why. Right now the habit is to say "yes" to AI wherever possible, and a reasoned "not here" or "not yet" is exactly what shows that a person understands how it works. Someone who really built it usually answers quickly and concretely. For example (hypothetically), a person who used to run the product at an online store might put it like this: Product descriptions went to AI right away, while search by order number stayed ordinary, without a model, because people needed exactly the number. Refunds stayed with people: on real requests AI confused similar cases, and a wrong refund is hard to take back. The chat assistant was removed, because it answered what was already on the product page, and every answer cost money. And every week someone on the team reads a share of the descriptions AI has written. After that, it's enough to ask "and how did you find that out?" to hear whether a real check stands behind the answer. And if the answer is "everywhere we could" or "we'll take a better model", there's most likely only vocabulary behind it so far. #EugeneOnProduct
1
24
A founder has closed a round at a high valuation, and when they tell me about it, they start with the valuation, while I usually wait for the moment they start talking about users. The valuation itself tells me little: it's the price at which the investors in this round agreed to buy a share, based on their understanding of the market and their expectations, and it doesn't always say what the company has built. What it does do is raise the bar: the next investor will most likely pay more only if the company has grown into that price. So my questions are almost always about what has changed since the previous round: 1. Who uses the product now. Whether people come back without reminders, how much the company earns and spends on each customer, and what works without the founders' daily heroics. 2. What you know now that you didn't know at the previous round. Not how many people you hired, but which doubts went away. 3. Where each jump in growth came from. And whether it will happen again next month if you stop "pushing" it. If there are answers, the valuation has a story behind it, and the next round continues that story. If there aren't, the number is, it seems, more of a debt than an achievement.
1
17
Let's talk about metrics for products where AI plays, if not the main role, then a significant one. Ask people whether they trust the assistant inside a product they use every day, and most will say "mostly, yes". Look at what they actually hand over to it, and the answer is often different. Take a hypothetical sales team with such an assistant in its CRM. In the first month it summarizes call notes, and the managers check every number. In the second month they let it draft the proposal to the client. By the third, nobody rereads what it wrote into the deal record. A survey rarely shows this, but behaviour shows it right away. So when AI does part of a product's work, I'd look at its metrics a little differently and take the following into account: 1. The first useful result. How fast a person gets it, what share of tasks actually ends done, and how much has to be fixed by hand. 2. Growing trust. Whether people come back to the same task and, over time, hand over something more important than before. 3. The effect on the business. What changed in money, time, the number of errors, or the quality of the work. Trust is probably measured least often (if you do measure it, tell me in the comments how), and it has a trap: corrections can drop because the product got better, or because people simply stopped checking. So what people hand over only means something next to whether the result got worse. And the strongest signal, it seems, is not that the assistant is used more often, but that it was trusted with something more important than a month ago, and the result didn't suffer.
2
31
I want to share one more thing I notice. When I work with founders as a mentor, they often come to me with a question about features: here are three, which one to build first, or here is a mockup, how should this step work. It's a fair question, and it's easy to spend the whole hour of the call on it. But behind this question there is usually something the founder doesn't know yet and the feature is supposed to find out, and nobody has said it out loud. So before discussing how a particular feature should work, I ask one thing: what don't we know that this feature should tell us? Once that is named in one sentence, the conversation changes. The next question is about how to see the answer in what people do, not in what they say. Sometimes it's enough to do the work by hand for a few customers, to ask for a real commitment (a paid pilot, time in their calendar), or to look at how people already work around the problem today. Building a feature isn't that expensive these days, but it stays in the product, people start relying on it, and removing it is harder than adding it. So we try to answer the questions first 👆 #EugeneOnProduct
1
2
49
When I work with founders and they show me their roadmap, I usually ask about one line: why is it here, and what would have to happen for it to move? Quite often there is no answer: the line has a name and a date, and nothing else. That’s exactly how a roadmap turns into a list of promises. Someone from sales sees the date and takes it to a customer, and a few weeks later, when new data says something else matters more, moving that line already means breaking a promise the product team never made. That’s why next to every line I recommend keeping the following: 1. What problem it solves 2. What the team knows about it right now 3. What result it expects Then, when the data changes, everyone can see why the order changed. It’s also very useful to agree in advance on what can move a line, who decides, and who is told and how quickly. And the lines that really were promised to someone (by a contract or a regulator’s deadline) are marked separately and don’t move quietly. So if the plan has to change tomorrow, each line can be explained in one sentence. #EugeneOnProduct
1
25
Imagine a team two weeks after launch: ten people have signed up, and the meeting is already about how to get a hundred. The launch post brought seven of them, two came from a thread in a professional community, and one came through an intro from someone's former colleague. So the plan writes itself: the same launch post, only bigger. Now look at what those ten did after the first day. Several of the seven looked around once and didn't come back. The one from the intro opens the product every Monday, because that is when he puts together a report for his team, and last week he loaded his own data and asked to add a colleague. That one person says more about what to scale than the other nine, because he came for a job that repeats in his week, not out of curiosity about something new. Ten users is too few to average and enough to know each one by name. So before more money goes anywhere, I'd write one row per person: where they really came from, what they were trying to get done, and whether they came back without a reminder or gave something of their own. And I'd scale only the channel and the words that brought the people who came back. #EugeneOnProduct
1
2
40
Working with different projects and founders, I noticed one thing: companies with two or more founders seem to survive more often than those with just one. And I don't think it's because two people have more hours in the day. A second founder gives things that are hard to see from the outside: their "no" is hard to wave away, they take on their own share of the responsibility, they remember what you decided and why, and they ask the question you'd rather not ask yourself. And one more thing I've seen more than once: when something drags on not for a week or two but for months, and there isn't money for everything, one of them starts to lose heart, the other notices and holds on for both, and some time later it's the other way round. Today it's easy for a solo founder to decide they don't need a second one, since everything can be discussed with AI. An agent really can take part of this work: the objection before a decision, the memory of decisions, the uncomfortable question. But only if you explicitly allow it to disagree, bring it in before the decision rather than just to write the text afterwards, and keep a record for it of what you've already decided. By default it will most likely agree, and sometimes it can even strengthen your confidence in a wrong step. The rest it can't take at all. No one can hold it accountable, it loses nothing if the company closes, and cheering you up costs it nothing. Want to try an experiment? Recall the last three decisions where AI changed your mind in an area where you're the expert. If there are none, you have a good assistant, not a second founder. #EugeneOnProduct #StartupAdvice
1
29
When a new person joins a team and doesn't yet know anything about the project and/or the company, they're almost always told two things: what needs to be done, and in which cases it's better to ask first. And the second part usually goes unnoticed — it comes up in passing, somewhere in the kitchen, in the smoking area, in small talk. When work is handed to AI, for some reason people focus only on what needs to be done, and even then — not always describing the actual result concretely. 🙂 A person can figure out much of the second part on their own, from experience, while AI knows exactly what it's been told, and if the goal is described vaguely and nothing is written about boundaries, it will most likely fill them in itself, the way it thinks is reasonable. That's why I suggest looking at handing work to AI roughly the same way as hiring a new employee onto the team: 1. Goal: what result you want to see. 2. Boundaries: which data it may work with and which must never be sent anywhere, what it may do on its own, and what only after a separate "yes": sending, publishing, buying. 3. A result you can check: something you can look at and correct, plus an honest note of what it couldn't verify. 4. Stopping: agree that if there's no safe way forward, AI stops at that exact step and asks, and carries on with the rest of the work. Without any one of these, AI may simply start doing something other than what you actually want from it. #EugeneOnProduct
1
36
When I meet founders as a mentor, I often pay attention to one word: "We". In the pitch it's everywhere: "We" found the problem, "We" built the first version, "We" see the market. It sounds like one team, and at a first meeting that's usually how it's presented. But a few months later, with some founders, the word "We" changes. What worked sometimes becomes "I": "I" pushed for this, "I" found that client. What didn't work becomes "He" or "She": the release slipped because of "His" part, profit dropped because of "Him". And it isn't about manners — in situations like these, it seems the founders never agreed who owns which decisions, so wins and losses have no "owner" and get assigned after the fact. In a team that is really together, "We" survives even the story of a mistake: "We" were sure about this, the data said otherwise, "We" changed it. If "We" only shows up in the pitch, that's probably the only place the team exists. #EugeneOnProduct
1
27
1/5 "Engineers own features end to end now. Why are we still paying for product managers, designers and engineering managers?" To some, the question may seem valid if the goal is a line in the budget. #EugeneOnProduct #ProductLeadership
1
1
45
4/5 If the roles are merged and, say, nobody is explicitly responsible for the three things above (and there may be more), the process got worse, even if every release got faster. In a few months some feature will have no owner, and it will show.
1
1
8
5/5 Merging roles makes sense when it makes the process better, not only cheaper. And if a company, whatever its size, isn't ready to rebuild its processes right now (which is fine), it's better not to chase the trend for the new role at all.
1
6
When someone shows me a prototype these days, the screens are far from the first thing I look at: with AI, good-looking screens take a weekend (I build mine the same way myself). What I'm waiting for is what they say about users. Usually one flow gets shown, and its author drives it like a familiar road: types the example that's sure to work and steers around the rough spots. Nothing wrong with that, a demo should work that way, and some things may not work at all yet, as long as that's a decision. What worries me is when nobody mentions that this is one flow out of several, why this one was picked (say, because it's the most common), and what the users they talked to said or did with it. Because whatever gets steered around in the demo doesn't disappear in the product: someone else's data that looks nothing like the example, errors, who is allowed to see and change what, and a person who fixes it all when it breaks. The screens may well please whoever's in the room, an investor, a CPO or a department head, depending on the meeting. So after a demo I ask the same thing every time: what did we learn that's new, and what do we do next? #EugeneOnProduct
1
21
Everything I posted in September, grouped by topic. AI-native product A cheap AI feature nobody uses can live for months. Agree on one number at launch: teapawt.org/etsarkov/status/209620… Working on one narrow slice of support doesn't earn an AI agent autopilot for everyone: teapawt.org/etsarkov/status/210061… Four layers a competitor with the same model can't rebuild in a week: teapawt.org/etsarkov/status/210493… From 0 to 1 Three questions about first users to answer before writing code: teapawt.org/etsarkov/status/209735… Segment, then hypothesis, then questions, so interviews test something: teapawt.org/etsarkov/status/210128… Why product people say "it seems" so often: teapawt.org/etsarkov/status/210384… Product leadership in the AI era What to do with the NO in your head when you're handed a solution instead of a goal: teapawt.org/etsarkov/status/209481… Hiring a CTO when you can't check the code yourself: teapawt.org/etsarkov/status/209803… One person, several AI models, one computer: teapawt.org/etsarkov/status/209871… Making PM work visible as decisions, not busyness: teapawt.org/etsarkov/status/210242… Operator, advisor, and mentor What a monthly investor update should carry: teapawt.org/etsarkov/status/209549… Same advice, opposite effect at different stages: teapawt.org/etsarkov/status/209989… Mentor, advisor, introducer: three different jobs: teapawt.org/etsarkov/status/210316… #EugeneOnProduct
1/3 AI models get cheaper every month. An expensive AI feature used to remind you it was there, through the bill. A cheap one may never remind you at all. #AIProduct
1
15
The hardest part of an AI product to copy is most likely not in the product at all. A competitor with the same model can rebuild the product's screens and features in a week or a little more. What the team has built up together with its users can't be rebuilt that fast. I see four such layers of protection for a product: 1. The result users come for. People use the product for some result or process, and can name it without saying the word AI. 2. What the product knows about its users. Over time the product has "learned" their rules, history and exceptions, so it works better for them than on day one, while a competitor starts from zero. 3. Its place in their work. The product is built into how people work every day, and removing it means going back to doing part of the work by hand. 4. Proven effect. There's a result in numbers that users already show, on their own, to whoever they answer to: their managers, the board of directors, or their own investors. Each of these layers has to be "earned" again, and a competitor can't skip that. If a team can honestly show one or two, that's already good 🙂 #EugeneOnProduct #AIProduct
1
14