“Oh, Paris!”“Letters of the 20th Century”“Master and Moscow”

What the work was

From February 2025 to June 2026 I did research for the artist and type designer Yuri Gordon.

Four projects:

• the literary map “Oh, Paris!”;
• the map “The Master and Moscow”, following Bulgakov and Mandelstam;
• the infographic “The Catastrophe of the Silver Age”;
• research for chapters of the book “Letters of the 20th Century”.

A little background. By early 2025 the book had stalled. The last chapter had come out on 21 November 2024, and he opened it himself with the words “after a long break”: chapters had been appearing less and less often — there were two in the whole of 2024. After that, another six months of silence; the author admitted to me that the book was going hard and he had stopped writing. The book existed, and still exists, only as posts on Facebook* — lose access to the account and the book is gone.

I offered to preserve all the chapters and pictures — on one condition: he would carry on with the book. “At least you’ll write a couple more chapters.” In June we launched the AI project, and the book came back to life. Over the next year and a half more than twenty chapters came out — ten times as many as in the whole of 2024. I did the research for ten of them.

Before the project Yuri Gordon was deeply sceptical about neural networks. I spent hours showing them to him — on the phone, on Zoom, sharing my screen. Lectures and workshops for an audience of one. Later he said I had explained everything to him in literally three seconds. That is how AI entered Yuri Gordon’s projects.

The book took up the most space in my case study: Soviet copybooks, typesetting fonts, the lettering of the Moscow metro, shop signs and windows, book covers, monumental propaganda, type designers, the Thaw, the 1957 Festival and the 1980 Olympics. The research for ten chapters is mine.

Seven neural networks took part in the work, often in parallel on the same task: ChatGPT, Perplexity, Gemini, Claude, DeepSeek, Grok and Qwen. In total I handed the author 61 pieces of work.

The ChatGPT project was called “Digital YG”: we were building the author’s digital twin — his thoughts, his voice. A name for my half of the work appeared six months later, in his own chapter: “my partner and neurocaptain Julia Reprintseva”. That is also the title of the book that grew out of all this. More on that below.

There was a lot of material, of many kinds, scattered across seven places: 45 conversations in ChatGPT, 34 documents on Google Drive, 14 PDFs on Yandex Disk, two Telegram channels, 21 Facebook posts and folders on my computer. A single ChatGPT conversation held 71 messages — can you imagine? Rereading all of it, pulling out the dates and the briefs, working out which report belonged to which publication, checking the numbers — that would have taken months.

And while at the start the client was sceptical about AI, his attitude changed as the work went on. Yuri Gordon called the research for the chapter on typefaces “our best performance”. About the metro he noted that what Perplexity found in half an hour would have taken him two weeks. Plus the same again for sorting. About the last research I did for him: the best thing AI had done.

What makes this case study unique

I won’t retell every page. I’ll tell you about the most interesting parts.

Everything in one place. Sixteen months, four projects, seven sources — all structured, attractive and conveniently laid out. No more trying to remember what you did with the last year and a half of your life. Open the link — and voilà, it’s all in front of you. No long, painful hunt for a prompt or for the research it produced. What came out is a real map of the work. (Just as long as HR doesn’t start demanding portfolios like this from everyone. Don’t tell them. Or come to me — I’ll teach you to do the same.)

How the site works

Every row opens right inside the map: the brief, what was done, with which tools, what came out. If the short version isn’t enough and you want more, the row has its own page.

On the page everything is arranged in layers. At the top is the author’s brief with its date. Below it, what was done and with what. Next comes the brief itself — the actual document the model received, with the full prompt, which you can select and copy. Then the results: the PDF he published on his own pages, and the full texts folded into expandable blocks: click and it opens, don’t need it — close it, nothing to download. At the very bottom are the publications: the author’s posts about this work and my reports, in chronological order, each with a screenshot and a link to the original.

And everything is linked together. From a research project to the chapter written on it. From a chapter to the posts about it. From a report to the publication where the author posted it. From any page to the next: the thirteen research projects can be read in a row, like chapters.

The same material is shown in several ways, and each is a different pleasure. As a page — to study. As a deck of cards — 241 invented quotations, the quote on the front, the fact-check on the back. As a page-flip book — to read straight through. As an English version — to show a client abroad: the whole case study has been translated into English, from every research project to the page-flip book, the map and the games. On top of that, the reports are there as PDFs, exactly as the author received them.

Four projects — four products. The project pages describe each project in detail: the brief, the prompts, the reports. The “Artefacts” section on the main case page holds the things themselves. I came up with them along the way, while I was building the landing page. The quote book, the book, the infographic, the map — all in digital form, and you can do more than open and leaf through them. Each one lets you do something: flip a card, filter the poets, switch the route.

“Oh, Paris!” has a quote book: a deck of 241 cards, a fake quotation on the front and its exposure on the back. “Letters of the 20th Century” has a page-flip book. “The Catastrophe of the Silver Age” has an infographic: 25 poets and their fates before and after 1917 (I made it interactive and in the style of the whole portfolio). And for “The Master and Moscow” I made an interactive map, my own plan of 1930s Moscow. It has three layers: Ivan Bezdomny’s chase after Woland, Bulgakov’s addresses and Mandelstam’s Moscow. The route has a switch: “AI version / After checking”. Press it and you see how the neural network sent Ivan along the Garden Ring and Tverskaya to the Central Telegraph. In the novel he runs through Spiridonovka, the Nikitsky Gates and the Arbat lanes down to the Moscow River. Hover over a point and a quotation from the novel appears. And every map has a link to the original on the author’s page.

AI mistakes are not hidden. In the research for “The Master and Moscow” the neural networks made mistakes. They mixed up the lane where Ivan climbed into the river. They invented nine Mandelstam “quotations” — no such lines exist in any of his texts. And they gave Bulgakov’s line about “gingerbread towers” to Mandelstam. Meanwhile the fact-check, done by the same AI, said of the route: “No errors.” The author spotted the route error himself back then and called the Mandelstam section an AI failure. But the rest — about twenty of them — surfaced only now, when I was building the site and checked everything again. For the site I could have quietly fixed it all. But then it would no longer be the document the author received. So the text stayed as it was, and next to each mistake there is a note and the correct version. You can see both what the AI said and what turned out to be true. To my mind, that is the main thing about working with neural networks: don’t take them at their word.

Nothing is lost. Every document has a date and a link to the original source. The numbers aren’t made up: 245 sources, 111 links, 61 pieces of work, 13 research projects — all counted from the files, and any of them can be checked again.

The prompts are open. Usually people hide their method: “message me and I’ll tell you”. Here the brief is given in full — goal, inputs, quality criteria, output format, checks. Select, copy, run it yourself. You don’t need me for that.

You can see who did what. There is a separate table: the author, me, the AI — stage by stage. Yuri Gordon chose the topics; I set the tasks for the models and checked the sources.

The cherry on top: along the way an idea was born to make an e-book of my own. “Neurocaptain”: a cover, a title page, a table of contents, and it turns pages like a real book. Each spread has the opening of a chapter and the author’s words about the work on the left, and the research sheet on the right: what was asked, what was done, with which tools. Bookmarks, notes, search, three paper modes. And an English version. A case study about work unexpectedly turned out to be a product you can read as well. And that is a god-tier portfolio.

Then I thought: if all of this can be leafed through and flipped, why not play with it? That is how three games appeared in the portfolio. You don’t need to look for them in an app store: open this site on your phone, tap “Add to Home Screen” — and the game sits as an icon next to your other apps. I could have gone further and built an actual app — who knows, maybe I’ll get to that too.

“True or AI?” — you are shown a quotation, an address or a date from any of the projects and have to decide: is it true, or did the neural network make it up? Spoiler: it’s harder than it looks. Some of the fakes are so good I’d print them on posters. “Catch Woland” — Ivan Bezdomny’s chase across Moscow: at every crossroads a line from the novel and a choice of where to turn. Follow the AI version and you end up at the Central Telegraph, where Ivan never was. “The Editor’s Desk” — a neural-network report full of traps; you have to find all the mistakes. That one is closest to my actual job.

And if you can play, you can learn. This site also has a guide, “How to catch AI lying”: ten signs of a fake, each with an example from this project and what to do about it. How to notice that all the authors speak in one voice. How to check a book that may never have existed. And why you must not ask AI to check itself: mine once wrote “No errors” — in a report that had several.

And a brief builder. I put my briefs together using one scheme: goal, inputs, quality criteria, output format, checks. Now anyone can build a brief with it: you fill in the fields, and ready-made text for ChatGPT, Perplexity or Gemini appears on the right — with my checking rules. Press “Copy” and off you go. As an example there is a real brief about the lettering of the Moscow metro.

So a case study about work first became a book, then a game, and then a textbook as well. A portfolio you can play — I haven’t seen one like it anywhere yet.

* Facebook is owned by Meta Platforms Inc., designated an extremist organisation; its activities are banned in the Russian Federation.