Skip to content
missumai
Steal this idea· 8 min read

Steal this idea: Visualizee swapped its expert editor for a chat that writes the prompt

A solo developer's AI rendering tool sat at about $150 a month for two years behind a node editor. He replaced it with an assistant that asks what you want and writes the prompt itself. Here's the pattern, the numbers, and what to watch before you copy it.

monthly recurring revenue, Stripe-verified, from 282 active subscriptions (8 Nov 2026)
$7,899
all-time revenue, Stripe-verified (8 Nov 2026)
$96,232
a month before the change, and close to $10K a month after it, in the founder's words
~$150
of evenings to build the first, node-based version, founder-reported
5 months

Visualizee turns a sketch, a SketchUp screenshot or a photo into a photoreal render for architects and interior designers.⁠[11] It is built by Piotr Obidowski, a software developer in Poland with more than 20 years of experience, who told Starter Story he runs it alone next to a full-time job as a tech lead.⁠[1]

His story is a useful one to steal from because the thing that changed wasn't the audience or the idea. It was the way people talk to the product. The first version asked users to wire up nodes. The current one asks them what they want, then writes the prompt for them.

The pattern in one sentence

If your AI product makes people learn your tool before they get a result, put a chat in front of it that asks a few plain questions, writes the expert prompt and picks the settings itself, shows what it assumed, and keeps a precise manual mode for the people who want one.

What it has done so far

TrustMRR, which reads Visualizee's Stripe account through an API key, showed $7,899 in monthly recurring revenue from 282 active subscriptions, $8,108 in the last 30 days and $96,232 all-time on 8 November 2026. It lists the founding date as August 2023.⁠[2] That works out to roughly $28 per subscription a month, by our arithmetic. That is the verified part.

The before-and-after is Piotr's own account. He says he launched almost three years ago, was stuck at about $100 a month for the first two years, and that after the change revenue went from about $150 to nearly $10,000 a month, with one-time payments added on top of the MRR.⁠[1] We can't see the revenue history ourselves, so treat the "before" number as founder-reported.

  1. Aug 2023

    Founding date on TrustMRR.⁠[2]
  2. Until Sep 2025

    The homepage offers three free renders and says sketches, images, 3D models or text prompts become renders. No assistant is mentioned.⁠[12]
  3. 9 Oct 2025

    New homepage: "Upload or describe", "without the software learning curve".⁠[13]
  4. 3 Nov 2025

    First snapshot we found with "Chat with Vizzy": describe your vision and "Vizzy helps you craft the perfect prompt".⁠[14]
  5. 8 Jan 2026

    Blog post introducing Vizzy as the centrepiece of the v3 release.⁠[3]
  6. 24 Mar 2026

    "Vizzy goes agentic": the three workflow cards disappear and Vizzy picks the workflow itself.⁠[5]
  7. 2 Sep 2026

    Starter Story interview: about $150 a month before, close to $10K a month after (founder-reported).⁠[1]
  8. 8 Nov 2026

    TrustMRR: $7,899 MRR, 282 active subscriptions, $96,232 all-time.⁠[2]

What was wrong with version one

Piotr says the first version took about five months of evenings and was a node-based system similar to ComfyUI. Users told him directly it was too hard to understand.⁠[1] In the interview he walks through it: upload an image to make one node, connect it to another, write something, click generate, with "a lot of different parameters around".⁠[1]

“For two years I blamed slow growth and marketing when the real problem was that my product was too hard to use.”
Piotr Obidowski, on Starter Story⁠[1]

That is the problem this pattern solves. A node graph is great for people who already think in nodes. An architect who wants a client picture by Thursday doesn't, and shouldn't have to.

What replaced it

A chat that does the prompting. Piotr describes the new layer as an assistant that helps people first understand what they want, asks follow-up questions, and then builds the prompt and sends it to the image services.⁠[1] Visualizee's own blog puts it bluntly: most AI tools make you "a prompt engineer first and a designer second".⁠[3]

It speaks the user's language, not the model's. The docs contrast "4k, photorealistic, octane, ultra detailed" with a sentence describing the result (a dusk shot of this house, interior lights on, a wet driveway) and say Vizzy turns that into the model, the settings and the detailed prompt.⁠[7] A February post says they chose not to build a keyword optimizer: "If your natural language is good enough for a design review, it's good enough for Vizzy."⁠[4]

Then the menus went too. In March, the blog says, users still had to pick one of three workflow cards (text-to-image, image-to-image, video) before typing. That update removed the picker: Vizzy reads the message, the attached images and the conversation, and chooses the workflow and settings itself.⁠[5]

You type, no imageVizzy makes a new concept from text.⁠[6]
Sketch + “make it photorealistic”A render that keeps your design.⁠[6]
“Animate this building”A video from the image.⁠[6]
12 photos + one changeA batch, one render per photo.⁠[6]

It stays honest about what it decided. Under each result, the docs say, Vizzy lists what it assumed for anything you didn't say. Settings you set by hand win over its defaults. A "Plan first" mode makes it ask and show a plan before rendering, a visible setting change waits for you to approve or deny it, and when it can't tell which image is the reference it asks instead of guessing.⁠[6]

The expert knowledge moved inside. Instead of asking users for long prompts, Vizzy picks a specialist (architecture, product photography and so on) for each message and loads that job's "skills"; power users can type / to add up to five themselves.⁠[8]

Two other changes happened at the same time

Be careful about giving the chat all the credit. Piotr says that with the pivot he also moved from one-time payments to monthly subscriptions, and his second lesson is to charge subscriptions from day one.⁠[1] Revenue that repeats every month grows differently from one-off sales, so part of the jump may come from billing, not the interface. He doesn't split the two, and we can't either.

The timing is also fuzzy. In the interview, published 2 September 2026, he dates the pivot to "six months ago".⁠[1] The homepage was already advertising Vizzy on 3 November 2025, and the workflow cards were removed on 24 March 2026.⁠[14]⁠[5] We don't know when the interview was recorded or which step he counts as "the pivot", so we'd read it as a series of changes from late 2025 to spring 2026.

The stack, as he described it

Built withCursor, and more recently Claude Code.⁠[1]
App and hostingNext.js on Vercel.⁠[1]⁠[2]
Database, sign-in, storageSupabase.⁠[1]
PaymentsStripe.⁠[1]
Image generationA third-party GPU service that runs the image models.⁠[1]
Growth and monitoringAhrefs for SEO, MailerLite for email, PostHog, Google Analytics and Sentry.⁠[1]
NewerA hosted MCP server, so Claude, Claude Code or Cursor can render with a Visualizee account.⁠[10]

How to run it yourself

The steps follow what Visualizee describes. Where a step is our own suggestion, it says so.

  1. Find where people stall before their first result

    Piotr's fix was "removing the thing standing between a visitor and their first success", and he says the data showed it: before, only a couple of sign-ups; after, people started rendering.⁠[1] Count how many new users get one finished result on day one.
  2. Write down the questions an expert would ask

    Lighting and time of day, materials, mood, and what to keep or change are the things Visualizee's docs say a good request names.⁠[7] Your niche has its own short list. That list is the assistant's script.
  3. Let the chat write the prompt, not the user

    Users describe the outcome; the assistant asks follow-ups and turns the answers into the model prompt and settings.⁠[1]⁠[7]
  4. Remove the menu in front of the chat

    Detect the job from the words and the attachments, and ask only when it is truly unclear.⁠[5]
  5. Show your assumptions and let people override them

    Assumptions listed under the result, manual settings that win, and approve or deny before a change.⁠[6] This is what keeps a "magic" chat from feeling like a black box.
  6. Package the expertise as skills

    One specialist per message, with named skills a power user can add by hand.⁠[8]
  7. Keep a precise mode for the people who want one

    Visualizee now also has Studio, a beta where you mark regions on one image and render changes "without a chat"; the docs pitch the chat for exploring and Studio for perfecting one image.⁠[9]
  8. Charge by the month

    His other lesson, in his words: charge subscriptions from day one.⁠[1] Today's plans are $20, $40 and $80 a month, sized in renders.⁠[11]
  9. Change one thing at a time (our suggestion)

    Visualizee changed the interface and the billing in the same period, which makes it hard to know which did what.⁠[1] If you can, ship the chat first and measure first results before you touch pricing.

Gotchas before you copy it

  • A chat on top of a model is easy to copy. In the interview, Pat Walls describes the simplified product as essentially a ChatGPT wrapper.⁠[1] What is harder to copy is the domain knowledge inside: the specialists and skills, and the rules about what to keep.⁠[8]
  • Chat isn't always faster. For one exact edit, Visualizee added a mark-the-area mode instead of more conversation.⁠[9]
  • Every turn costs you money. Plans are sized in credits and renders a month, with heavier settings using more.⁠[11] An assistant that renders several options per request needs limits that match your costs.
  • Don't hide the decisions. If the assistant changes settings silently, users can't fix what went wrong. Visualizee makes every visible setting change wait for approval.⁠[6]
  • Keep your own numbers straight. The prices in the interview ($35 or $80 a month), in the founder's TrustMRR text ($15, $35, $80) and on today's pricing page ($20, $40, $80) all differ.⁠[1]⁠[2]⁠[11] Prices move; date the ones you publish.

What to take from this

  1. 01

    Friction looks like a marketing problem

    He blamed marketing for two years.⁠[1]
  2. 02

    Ask, then write the prompt

    Follow-up questions, then the assistant builds the prompt.⁠[1]
  3. 03

    Speak the customer's language

    A design-review sentence, not a keyword list.⁠[4]
  4. 04

    Show what you assumed

    Assumptions under every result, approval for changes.⁠[6]
  5. 05

    Keep an expert mode

    Studio for one exact image.⁠[9]
  6. 06

    Bill monthly, and measure separately

    The billing changed alongside the chat.⁠[1]

The same idea shows up in a different form in FootballGPT's niche assistant, where a saved coaching profile does the prompting so coaches never retype their context.

All stories

The newsletter

More stories like this, by email

Occasional email with new projects on the leaderboard and new build stories. Every number sourced, like the site.

We'll email a link to confirm. Unsubscribe in one click.

Sources

  1. [1]I Failed For 700 Days Before My App Hit $10K/MonthStarter Story (Pat Walls interviews Piotr Obidowski) · Sep 2, 2026 · checked Nov 8, 2026We read the transcript printed on Starter Story's page. It looks machine-made (for example it spells the assistant “Visi”), so our quotes are short and the figures in it are the founder's own.
  2. [2]Visualizee.ai on TrustMRRTrustMRR (revenue verified with a Stripe API key) · checked Nov 8, 2026Figures as shown on 2026-11-08 (the page says last updated that day); they move daily. The value proposition, audience, user count and pricing on that page are the founder's own text.
  3. [3]Stop Prompting, Start Conversing: Meet VizzyVisualizee blog (Piotr Obidowski) · Jan 8, 2026 · checked Nov 8, 2026
  4. [4]Why We Built Vizzy to Think Like a DesignerVisualizee blog (Piotr Obidowski) · Feb 12, 2026 · checked Nov 8, 2026
  5. [5]Vizzy Goes Agentic: Your AI Design Assistant Now Thinks for ItselfVisualizee blog (Piotr Obidowski) · Mar 24, 2026 · checked Nov 8, 2026
  6. [6]Vizzy overviewVisualizee Docs · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  7. [7]Talking to VizzyVisualizee Docs · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  8. [8]SkillsVisualizee Docs · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  9. [9]StudioVisualizee Docs · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  10. [10]Visualizee MCP serverVisualizee Docs · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  11. [11]Pricing: Plans from $20 a Month, Free to TryVisualizee.ai · checked Nov 8, 2026Read on 2026-11-08. Prices, limits and features on a live product change often.
  12. [12]Visualizee.ai homepage, 17 September 2025Internet Archive (Wayback Machine) · Sep 17, 2025 · checked Nov 8, 2026Read through the Internet Archive on 2026-11-08. A homepage snapshot shows what the site said that day, not when the app itself changed. The same wording appears in snapshots back to July 2024.
  13. [13]Visualizee.ai homepage, 9 October 2025Internet Archive (Wayback Machine) · Oct 9, 2025 · checked Nov 8, 2026Read through the Internet Archive on 2026-11-08. A homepage snapshot shows what the site said that day, not when the app itself changed.
  14. [14]Visualizee.ai homepage, 3 November 2025Internet Archive (Wayback Machine) · Nov 3, 2025 · checked Nov 8, 2026Read through the Internet Archive on 2026-11-08. A homepage snapshot shows what the site said that day, not when the app itself changed. This is the earliest snapshot we found that mentions Vizzy.

Something wrong or out of date? Tell us and we'll fix it.