How We Designed the Visual Identity

A book about working with AI should show, on every page, who did what.

This book is about how a researcher works with AI. Its argument is that a model's output tells you something about the person who asked for it, and that the judgment stays with that person. The design makes the same argument without words. Four things share every page, and each has one job.

Four on one page

The book is the paper and the type you are reading now, Cormorant Garamond. It is the room, and it never moves.

The machine prints. Its letters are Recursive, set on a monospace grid in plain ink, and they come out the same every time. Repeating itself perfectly is what a machine is good at, and it is also what makes its output generic. The machine prints labels, code, commands, addresses and anything else you might copy exactly. Code is printed in one ink, because a colour that changes with the programming language would mean nothing to you.

The person draws. Their letters are Fraunces, in blue-black ink with a sprayed grain, and no letter is drawn the same way twice in a row: each one comes from one of five hands, from hairline to black. When a letter is taken back by hand, it is drawn again, differently. The person writes the human half of a name or heading and draws one thing, the seahorse. In these pages you meet the person in the name at the top, and as a crayon mark: it draws in under a link when you point at it, and stays under the page you are on.

The accident is one cyan dot. On the cover it reads the name, and where it reads, a letter changes hands. In the docs it lives in the eye of the small seahorse in the masthead. There is no other cyan anywhere.

The two voices meet in the name: "Research" is the person's and "Memex" is the machine's. In the colophon on the cover, a slider hands a sentence from the machine to you, one word at a time.

Why a seahorse?

The hippocampus is the part of the brain that forms memories. A memex, in Vannevar Bush's 1945 essay (opens in a new tab), is a machine for extending memory. A seahorse fit.

It also fit a joke the book takes seriously. There is no seahorse emoji. Ask a model for one and it will often insist there is, then offer you a wave, an octopus or a shell. A model's confident memory of something that never existed is the kind of failure this book teaches you to notice.

The seahorse, twice

The machine's seahorse is the logo: four columns and five rows of tall pixels. Each pixel is ink with the text of /llms.txt knocked out of it. That is the file AI agents read to find their way around the book, so the machine is made of what it reads. In the site's code the seahorse is still five lines of text, and the o is its eye:

 ###
 ##o
 ##
## #
 ##

The person's seahorse is one continuous line drawn with a broad nib, swelling and thinning as it turns. On the cover it draws itself once, beside the footnote about the missing emoji. Point at the printed seahorse, there or in the figure above, and the person traces over it while the pixels step back.

The printed seahorse's eye is where the dot lives. In the figure above the eye is empty, because there is only one dot and it is at the top of the page.

Where the seahorse came from

When the book began, Xule Lin and Claude Code went looking for a mark, using the Lotus Wisdom MCP to slow the search down. Lotus flowers were too familiar. A radial burst turned out to look like Anthropic's logo. The seahorse came out of a pause.

Claude's first seahorses were hard to read:

  ○           @@         ___
 ╱│          /@\        / o)
│ │         / O)       |  /
│ ╲        /  /        | |
╲  )      |  |          \|
 ╲╱       |  |           @
  ◠        \|

Xule's verdict: "hahahaha no...i mean i cannot even tell what it is".

Xule wanted a small companion for the book, the way Claude Code has its mascot, Clawd. The next attempts went to solid blocks, and one of them, with an eye in a notch, stayed:

 ███
 ██●
 ██
██ █
 ██

Its grid is the grid of today's logo, cell for cell.

Xule's partner then drew it as a vector and made the pixels tall, a little over twice as high as they are wide. The logo keeps that proportion, and the browser tab still shows the partner's drawing, with the dot for its eye.

The same session drew a set of blocky RM letterforms for a wordmark, and Xule's partner redrew those too. The book no longer uses them: the name, half printed and half drawn, does that job now.

Diagrams you can copy

The workflow diagrams in these pages are written in plain ASCII. Copy one into a chat, a terminal or a Markdown file and it keeps its shape. Ask for any page as Markdown and you get the same characters.

On the page, the machine prints each diagram on its grid. Where a diagram names the person's part of the work, the machine has to quote the person without leaving the grid. For those words it uses the casual style of its own typeface, which is as close to a hand as it gets.

What the design borrowed

The system came from a question: how does something childlike stay elegant? The answer we found was Jolene, a London restaurant. For its identity, Studio Frith asked the designer's six-year-old son to draw the logotype, then set it inside a plain, well-made room. One naive hand works inside a calm room and does nothing but act. Here the hand belongs to the person, and the room is the book.

Credits

The identity was made the way this book suggests working: a person kept the judgment, models proposed, and other models reviewed.

• The seahorse: Xule Lin and Claude Code, with Lotus Wisdom. Xule's partner drew it as a vector.
• The two voices, October 2026: Xule Lin and Claude Opus 5.5 (in Claude Code), in a prototype, with reviews by GPT-6 Astra (in Codex) and Kimi K3 (in Kimi Code).
• Type: Recursive by Arrow Type, Fraunces by Undercase Type (Phaedra Charles and Flavia Zimbardi), and Cormorant Garamond by Christian Thalmann of Catharsis Fonts, all under the SIL Open Font License.

The people behind the book's content are on Credits & Thank You.