
Jurij Tokarski
Discovery Kept Its Prompts, Changed Its Interface
It began as one 15-minute AI interview, then became eight chats. Now it is nine one-click tools. The prompts survived all three versions.
A visitor asked about privacy.
Fair question. The AI chats on my site kept what visitors typed. Eight of them were Discovery, my set of AI tools for software founders.
My answer was its third version, with no chat. The prompts at its core, the instructions the model follows for each result, carried through all three.

One Interview Asked for Everything
Discovery began in 2025 as one 15-minute AI interview. A founder answered questions in a fixed order and got one PDF: business model, market, personas, competitors, scope, stack and budget. The core prompts were there from the start.
Most founders stopped three or four exchanges in, at a question they couldn't answer yet. Weeks of better prompts, shorter flows and smarter branching didn't change that. The prompts weren't the problem. The fixed order was.
Eight Chats Kept the Conversation
Early in 2026 I wrote that nobody finishes a 15-minute AI interview. Then I split it into eight tools, one deliverable each. A founder looking for a lean canvas wants the canvas, not an interview. The tools chained, but each one worked alone.
The split made each prompt standalone and extended it around the same core. My plan was a shorter interview, with some of my own knowledge built in so the tools would argue back.
So each tool was still a chat. Chat is still the native format for AI products, and the first version had been one too.
Why the Chats Had to Go
The visitor's question made me look at every place on the site that kept free text from strangers. Each one was a privacy risk.
The 404 chat became a one-turn AI terminal. The model sees only the broken address, and nothing a visitor types leaves the browser. The chat on my principles page went too. That left Discovery as the last chat standing.
At first I thought it only needed another iteration. Then the code gave me one more reason to drop the chat. About 2,900 of the 12,200 lines in Discovery existed only for the chat. The renderers and the PDF already worked on finished results. Founders came for the result, and one who wanted a lean canvas still had to talk their way to it.
Nine Tools Make One Call Each
So my brief for the third version had three steps: the user provides data, one AI call, a formatted and structured result.
One call can be checked. It returns one structured answer, and code can test, cap and price it before anyone sees it. A chat's answer depends on every turn before it.
The prompt core carried over from the chats; only the way I call the model changed. The third version went live two days later, on October 7, with nine standalone tools: the eight, plus a new Idea Generator.
The hub: nine tools on a numbered path, the Idea Generator first.
Each tool starts with one short form: every question on one screen, one run button.
Lean Canvas: five questions on one form, starting with the idea in one sentence.
This is the part the visitor asked about. Answers go to the model for the run. Nothing a founder types is stored.
The run is one Claude Sonnet call that returns structured output.
One call while the founder waits: answers, canvas, check.
Code renders that output as a one-page result, the same on screen and in the PDF, with a Markdown copy.
The full result: a verdict first, then where the idea is weakest, then the canvas itself.
Code also guards what comes back. If the model adds a sales line, like my name or "hire me", the run fails. The founder gets the tool's made-up example instead.
The Test Was Wrong, Not the Tools
I switched the tools on to see one thing: whether the flow held from form to result. The switch changed no prompt and no app logic. Caps in code bounded what that could cost: $3 a day, $20 a month and 30 runs per visitor a day.
The guards needed proof too, so every tool has an end-to-end test. It opens the tool in a real Chrome, fills the form like a founder, runs it and judges the result. It screenshots every step and downloads the PDF, JSON and Markdown. Then it checks what the model wrote: length limits, quotes that match the founder's own words, banned phrases, English only.
First, a run on the tools' made-up examples caught a real fault of mine. The Lean Canvas example failed its own check: two guessed numbers had no estimate mark. I marked them the same day.
Then the test ran against real model calls and failed five of nine tools, though every tool had returned a result. My test was wrong in two ways. It capped one field at 300 characters where the engine allows 600. It also checked the order of the weak points, which the page sorts by itself.
The test stayed as it was for three days. Then I ran all nine again, around one made-up idea: a booking link for mobile dog groomers. The same two checks failed, so I fixed the test and ran it once more. No check failed. Across both runs, a result took 9 to 24 seconds and cost about 4 cents.
One idea through all nine tools: nine one-page PDFs from real model calls.
Nine live runs, no failed check.
The loading and result screenshots above come from those runs too.
The Prompts Survived Three Versions
The interview asked for everything at once. The chats asked for one deliverable at a time, but turn by turn. A privacy question took the conversation out.
Now a founder fills one short form, and one click gets the result.
Three versions in, the prompts are the part that lasted. Only the interface changed.
Try it: varstatt.com/discovery
Got thoughts on this? Reply by email.

Jurij Tokarski
I run a solo business and produce software. Usually I'm shipping for clients or building for myself. Sometimes I pause to write about it.