The roast. The number. The pivot.
An honest answer to whether a three month build should exist, argued by four agents at once, with the cost of winning a single customer put on the table beside it.
What came out of it. He arrived with Frontier, about seventy thousand lines written by prompting over three months, an application that walks a small company through the EU's sustainability report. He did not want help finishing it. He wanted to know whether anybody would pay for it. Four agents went at it in parallel, the same report was filled in from two prompts and no app at all, and the answer that landed was not about the code: winning one customer on German search ads costs around 2,500 dollars, which is two and a half years of that customer paying. He left with the idea rethought rather than finished, and a different thing to sell.
The session
The full session · 1h 58m · Watch on YouTube
No prompt log for this session.
The hero of this one
Denes Kovacs
Senior Project Manager at EPAM Systems, Zurich
He likes building startup ideas, and he came on to have one roasted rather than helped. Two hours later the roast had landed, and the thing he had not costed was what it takes to win a single customer.
Worth stealing
Split a roast into one agent per question
The idea and the business model, the code and whether it can hold client data, whether there is an easier way, and whether anybody is searching for it. Four answers you can read against each other are worth more than one balanced paragraph that offends nobody.
Roast the roast
"You never calculated how much marketing money this is going to cost me, and that is why I am angry with you, because you did not point that out." The first review was polite about the part that mattered most. Asking again with the omission named is what produced the number.
Try the two prompt version before you defend the three month one
The same report, no app: research the space, say what information you need, and start a second agent to check the first. It got far enough to be a fair comparison, which is the only way to find out what the build is actually adding.
If anybody can ask for it, it is not an advantage
A public document plus a public model is a thing a thousand other people can also ask for. The question is what you know, or can reach, that they cannot.
Pick who you want to work with before you pick what to build
In your own time, building for people you do not enjoy is a job you have given yourself. It is a business question that gets asked last and decides most of it.
The last mile is a script, not a chat
Filling in the official spreadsheet was not the model typing values back at you. It wrote a Python script and put 125 of them into the file directly, which is the part a chat window cannot do.
What happened, when
Every line opens the session at that moment.
- 0:00 Old friends, and an MBA in sustainability
- 3:09 The wish: roast the idea, and the business model I do not have
- 7:24 Local models, and what you actually give up
- 14:30 Before the app: what happens if you just ask?
- 21:32 A small furniture company in Hungary, and a report from nothing
- 25:18 If anybody can ask for it, it is not an advantage
- 26:35 His own answer: the only moat is speed
- 33:17 No markdown reports on this show
- 34:19 Four agents, one for each thing worth arguing about
- 38:05 Can the build actually hold client data?
- 38:47 Is anybody searching for this at all?
- 44:51 A form built in a minute, to see how far a stranger gets
- 52:54 "I have no clue what question you are asking me"
- 55:25 A real company, picked off a search, filled in from public records
- 56:43 What an AI expert is usually selling
- 1:02:51 The roast comes back: pivot needed
- 1:04:33 So how would anybody find it?
- 1:06:19 The code review: a demo, and a key every visitor can read
- 1:09:50 Which people do you want to work with?
- 1:18:21 Roasting the roast: you never costed the marketing
- 1:29:48 The business case, and a market that was never the problem
- 1:32:22 2,500 dollars a customer, and two and a half years to earn it back
- 1:33:49 The question nobody else would have asked
- 1:39:09 "The advertising part, I completely missed out"
- 1:45:06 125 cells written into EFRAG's official template
- 1:55:18 Sell the help, not the app
- 1:56:50 What the next session is for
Show notes
He came on to be argued with. Three months and about seventy thousand lines of prompting had produced Frontier, which takes the Excel sheet a small European company is handed when a bank asks for its sustainability report and turns it into ordinary conversations, one topic at a time, in the company’s own language. The wish was not to finish it. It was to find out whether it should exist.
So the session ran two experiments side by side. Four agents pulled the thing apart, one each on the idea, the code, the easier route and the demand. And in another terminal, the same report got filled in for a real Hungarian manufacturer with two prompts and no application at all, ending with 125 values written straight into the official template by a script it wrote for the job.
The verdict that mattered was not about the software. Asked, on the second attempt and only after being told it had skipped the question, what it costs to find one paying customer, it came back with about 2,500 dollars on German search ads against a subscription that takes two and a half years to earn that back. His own words, at 1:39: “The advertising part, I completely missed out, honestly.” The thing he left with was a different offer, which is to sell the prompting rather than the app.
Last updated 15 September 2026. Published as-is: read what a prompt does before you run it. The disclaimer has the rest.
