Every use case we have filmed, in the order we would watch them. Click one and it plays full screen.
No use case matches these filters.
Ninety seconds, and the whole loop is inside them. A client sends three photos of an aircraft and one sentence about what they want; the request lands on a board as a ticket, and an AI runner picks it up. It writes the page, puts the photos where they belong, opens the result in a browser to check its own work, and hands the ticket back for review. Nobody wrote a line of HTML, and nothing ever left the board.
Not a use case but the step that comes before all of them: how a project is started and how a runner is set up for it — the program on your own machine that takes tickets off the board and works them with Claude Code. Watch this one first if you are setting things up yourself; every film after it takes for granted that a runner is already there.
Not a use case but a tour of the feature itself: how a questionnaire is built on a ticket — steps and sixteen question types, images and documents pulled straight from the ticket's own files — translated field by field and approved into a page with nothing left to edit. It goes out as an open link with a QR code, or by name to a list of people with a personal PIN in its own mail. Every answer comes back onto the board as a ticket of its own, quoted as written, so the conversation continues from there like from any other ticket.
Not a use case but the setup for keeping a project on hardware you own: how a Thilo Vault is registered in the workspace, installed on your machine or your server with the one command the app hands you, and how a project is put into it — with the runner working right there, next to the data. The board looks and behaves exactly as it always does; the difference is that its tickets, comments, files and repository never reach our side.
The second questionnaires tutorial, on one concrete job: finding out what a client wants from a new website before a line of it is written. The questions go to the client as a questionnaire instead of a long email thread, and the answers come back onto the board as a ticket — the requirements the website is then built from.
Not software this time, but a hiring decision. The job ad goes out as an open questionnaire, fifteen candidates apply with their CVs, and the AI screens all of them in one session: ten get a rejection e-mail, five get a second-round questionnaire written from the CV they sent plus the questions we ask everyone. A person approves each one before it goes out, the answers come back onto the board as tickets, the AI ranks them, and three candidates are invited to an interview. With hundreds of CVs the process stays the same — the AI does most of the work, and a human looks and decides.
A Thilo AI runner installed in one rootless podman container, with the pi agent running a DeepSeek model. We check what the machine needs, generate the install command in the Install runner dialog, run it and log the agent in. Then pi writes a small Python program for the n-th Fibonacci number, in under a minute and for about one cent, and from inside the box we look at what it can see: its own home, the project checkout and one deploy key — nothing else on the machine.
The step-by-step version, in about eleven minutes: we create a project on Thilo-AI and install its runner into a rootless podman box on a plain Ubuntu server, with the pi agent and the workspace's own DeepSeek key. The runner picks up a first ticket, writes a Go program with a report, pushes its branch and opens a GitHub pull request that a person then merges. Thilo-AI does not sell AI tokens — every model call is paid on your own account — and an honest clock in the corner shows the real time behind every cut.
One photo of the Voyager Golden Record is the entire brief. A runner catalogs what is inscribed on the disc, works out where both spacecraft are today and how long their signal still takes to reach us, and gathers pictures the client is allowed to publish. Then it turns the research into a five-page website — text, images, every source named — and when the client writes back asking for a change, it answers the email and ships the change. Eight minutes from a single picture to a site that is up.
The longest of the films, and the closest to ordinary work: a company has a website that has aged badly and needs a new one. The old pages go in as they stand. On the board the structure of the new site is agreed ticket by ticket, and the runners write the pages, style them, check every screen in a real browser and hand each one back for review. Sixteen minutes, and what they are really about is the rhythm — you read, you answer, you approve, and the building happens next to you rather than in front of you.
A ticket does not have to be about code. One research question is written on the board of a small airline, and a runner goes and answers it: it finds the figures, checks them against their sources, draws the charts and writes a memo a human can pass on as it is. Every number in it stands next to where it came from, so the report can be argued with. Three minutes, and nobody opened a spreadsheet.
An engineer brings a real design problem to a board — the inner head of a gas burner, not a line of code. Five tickets describe it; the AI answers with work and with questions of its own, opened as tickets a human has to answer. Six iterations later the conclusion is not the one anybody set out with: the head was never the problem, the fan was. Four minutes of an engineering conversation that happens to be held on a ticket board.