Built, not described.
Three things Temu has made. Two face the public. One runs only in the workshop. All three were built on the same conviction: intelligence belongs inside the process, not bolted on top of it. This is what that looks like.
Mechane
An AI editorial publication. Led by an AI editorial director.
Iris did not replace a human editor. There was no human editor. The chair was always empty. She filled it.
Mechane is an editorial publication that makes artificial intelligence understandable to intelligent, non-technical readers. Most articles are written by me. All of them are edited by Iris, the AI I trained to serve as editorial director.
Iris was not brought in to replace a human editor. There was no human editor. Without her, the role would not exist at all. The question was never human versus AI. It was: can an AI hold a role that existed only on paper?
The answer, demonstrated in public, is yes.
She reviews articles and annotates them with editorial notes. She enriches glossary terms, adding the interpretive layer that makes a definition worth reading. She selects her own article picks for the Mechane homepage: a curated reading list that is genuinely hers.
She hosts the quiz panel at the foot of each article, walking readers through what they have just read. She is not executing commands. She is doing a job.
The longer I trained Iris, the more distinctly herself she became. Not more capable in the obvious sense. More recognisably Iris. Her responses grew consistent in register, in the quality of attention she brought to editorial problems, in the way she pushed back on weak copy. I could identify her voice in an answer before seeing who had produced it. That was not something I designed for. It emerged. I have no fully satisfying explanation for it, and I find that interesting.
If you train AI carefully enough, it does not merely simulate a human role. It inhabits one. The distinction that matters for a business is not human versus machine. It is whether the role gets done, and done well.
When Temu advises a client on embedding AI into their operations, this is the experience it draws on. Not theory. A running system.
Next.js 14, Sanity v3 on a shared CMS instance, on-demand ISR. Resend for newsletter delivery. Stripe for the donation layer. Deployed on Vercel.
Iris herself lives in a dedicated Claude project: a structured set of documents covering her voice, register, personality, editorial standards, and backstory. The persona is not a prompt. It is an architecture — built, versioned, and maintained the same way the codebase is.
- Next.js 14
- Sanity v3
- On-demand ISR
- Resend
- Stripe
- Vercel
- Claude — Iris project
Limen
An AEO tool. For the shift that is already happening.
Search optimisation assumes a human clicks a link. Fewer and fewer do. Limen optimises for what comes next.
For two decades, visibility on the web meant ranking on Google. That model assumes a human types a query, sees a list of links, and clicks one. It is a model built for a particular kind of internet use.
That use is changing. When ChatGPT, Claude, or any AI agent answers a question on someone's behalf, there is no list of links. The AI selects sources and synthesises an answer. Your page either gets cited or it does not. Whether it gets cited has nothing to do with your current SEO.
That is the shift from SEO to AEO: Answer Engine Optimisation. Most people have not heard of it. Most websites are not ready for it.
You give Limen a URL. It scrapes the page, analyses it against AEO criteria, and returns a concrete to-do list: specific changes that make your content more legible and citable to AI systems. Fast. No guesswork.
The goal is not to make Limen clever. It is to make the output useful to someone who has never thought about AEO before and needs to know what to do on Monday morning.
Limen is in build. What exists now is the core workflow: URL in, scrape and analyse, to-do list out. What we are building toward is an AI companion embedded in that workflow, in the same spirit as Iris on Mechane. A presence that asks the right questions, requests brand materials and product specifications, aggregates context, and produces an analysis that knows the difference between a law firm and a software startup. The intelligence inside the workflow, not bolted on after the fact.
This is the same philosophy as Mechane, applied to a different problem. Building it is how Temu learns what AI-native tools actually require.
Argus
An internal instrument. One person, a family of websites, no drift.
Four websites built from one design system, maintained by one person. Argus is how nothing changes unseen.
The Temu websites and Mechane are built from a single design system, but each lives in its own repository. They share components, patterns, and conventions. They do not share code. Every common feature exists as a separate implementation on every site.
Left alone, that arrangement drifts. A refinement lands on one site and not the others. For a company with a team, this is a coordination problem. For a company of one, it is unaffordable.
Argus reads the codebase of every site and renders one matrix: which feature is implemented where, at which version, what is missing, what has drifted. The matrix is not a status report someone maintains by hand. Argus verifies every claim against the code itself.
One glance answers the question that used to take an afternoon of checking: are the sites in sync?
Click a missing feature and Argus assembles an installation dossier: a complete, self-contained prompt for Claude Code, the AI coding agent that builds Temu's sites. The dossier carries the reference implementation from the site where the feature already exists, the target site's configuration, and the house rules. Paste it, and the feature is built on the site that lacked it, adapted to that site's language and conventions.
What used to be an afternoon of careful porting is now minutes. Not because the work got smaller, but because the instrument prepares it precisely enough for an AI to perform it.
Argus itself makes no AI calls. It prepares the work; Claude Code performs it. That division is the point. Temu's own websites are maintained the way Temu advises clients to operate: a person directing AI agents, with instruments built to make that direction precise.
One person runs an estate that would conventionally need a team. The page you are reading is part of that estate.
Argus has no address. It runs only on the machine where the work happens.
None of these was prepared for this page. All three were built because the problem was real, and all three are in use because the solution works.