Every serious brand has guidelines, usually a PDF: carefully typeset, thorough, and opened by almost nobody after launch week. There is now another audience that is unlikely ever to see it. When someone asks an AI assistant about your company, it is unlikely to consult the guidelines at all. It searches the public information it can find, often starting with your website, and makes its best guess. Your positioning, your tone, and your services are already being paraphrased by systems you have never briefed.
We think that audience deserves the same rigour as any other. So we started with ourselves.
What we shipped
The Bodkin website now publishes a small, structured kit for machines, sitting quietly alongside the pages people see.
- llms.txt: an index of who we are, what we do, our work, and our writing, published using an emerging convention that makes a site’s important content easier for agents to find.
- capabilities.json: every service we offer, with capabilities and honest answers to common questions, as structured data.
- work.json: our case studies, clients, and disciplines.
- studio.json: the facts. Where we are, how to reach us.
- voice.md: our verbal identity, written as instructions an agent can follow when writing as Bodkin.
None of it is hidden. You can open every file in a browser, and that openness is the point: a machine layer that only one company can read is not an identity, it is a liability.
No assistant is guaranteed to discover or prioritise these files yet, and publishing them is not a switch that controls every answer. What it changes is what happens when an agent does look for authoritative information, or is pointed at it: there is something deliberate, current, and in our own words for it to find.
The interesting part is not the files
Anyone can upload a text file. The part we care about is where the content comes from, because nothing in the kit is written by hand and left to rot. Every file is generated at build time from the same data that renders the website itself, so publishing a case study adds it to the kit, and retiring a service removes it. The machine-readable layer cannot quietly drift away from the website, because both are made from the same source.
That is the difference between documentation and a system. A PDF is a photograph of a brand on the day it shipped. This is a live feed.
There is a small proof of the idea in this very post. Before it was published, it was checked against our own voice.md: UK English, no em dashes, the Oxford comma, and paragraphs written in the register the piece calls for. The same file that briefs an agent briefed this writing, and it has already earned its first revision. Writing our long-form essays against it exposed a paragraph rule that was too blunt, made for scannable site copy and wrong for sustained reading, so the voice now carries two registers. Guidelines that get used get better. That was always the theory, and it took one week to prove it on ourselves.
The dictionary half of an older idea
We wrote recently that the next identity will be recognised, not scanned. The argument was that a pattern only has to say who you are, because once a machine knows who it is looking at, the brand’s own domain can supply everything verified and current, in the brand’s own words.
This kit is an early piece of that second half, not the whole of it. A full dictionary needs more than facts and voice: signed words, proof of what the brand does and does not authorise, and answers a machine can verify rather than simply read. The pattern research continues in the open. The dictionary has its first pages.
Try it yourself
Open a fresh conversation with any assistant that can browse, ask it to fetch bodkin.studio/llms.txt, and have it describe the studio. Then ask it to read voice.md and write a sentence as Bodkin. If it comes back with our actual services, our real work, and a sentence we would happily publish, the system is doing its job. That is the test: not whether the files exist, but whether a machine that has never met us can represent us properly.
Where this goes
We are now building brand systems this way for clients: designed for both audiences from the start, with the human guidelines and the machine kit generated from one source of truth. We have come to call them AI-native brand systems, and we suspect that within a few years the term may sound as redundant as responsive web design does now.
If your brand guidelines are a PDF and the agents describing your company are guessing, we should talk. james@bodkin.studio
![Colophon spread from the book Ritual Technology, set entirely in Adrian Frutiger's OCR-B typeface. Columns of green, red, and blue monospaced text declare "this book is readable by both human(s) [and] machine[s]" beside a full character set, ASCII-art figures, and duotone photographs treated as data entries. A teal-tinted cat labelled GARY [RIP], a man and an upside-down dog in orange labelled TIM and LOULOU, and a blue-tinted portrait labelled DINA.](/_astro/hero.D6lxdkQ6_ZJvxUm.webp)
