AI Kitcraft, 21–22 September 2026

Last week the SIG ran AI Kitcraft, a hands-on AI tooling workshop at the 2026 Protocol Symposium, facilitated by Rafa and Sachin. It started from one premise: AI is in its Kit phase. Like every general-purpose technology before it, the real work happens in people's own local experiments long before it turns into firm-scale products, and the missing productivity everyone is waiting for is a sign of that stage. The workshop asked the next question: everyone uses AI their own way, so what happens when we need to work together?

Over four one-hour sessions, participants worked through Kit → Factory → Bridge, each with their own AI assistant working against one shared, public repository:

• Kit: list the AI setups you already use, and read everyone else's.
• Factory: make one of yours runnable by someone else, or by their assistant.
• Bridge: get one assistant to run several people's kits together, reliably.

What happened. Sixteen people joined the first session, fourteen of them participants. The workshop's own evidence pack shows the funnel narrowing at every stage: almost everyone wrote an inventory, and nine of twelve with a folder in the repo also wrote about how their setup relates to someone else's. That is the clearest sign people read each other's work. Just over half did factory work beyond the template, a quarter opened a bridge, and two wrote a show-and-tell. About seven people were still there in the last session, and each had built something of their own.

What worked. The framing, which focused on understanding the technology's current phase rather than on building a tool. The examples of what people had actually built. Working through a single shared repository, with a context tank, glossary, and resources that participants could ask questions of.

What didn't. Discord and GitHub were gates for people who don't already live in them. Expertise ranged from first-time GitHub users to experienced programmers. Four sessions in two days, in a crowded symposium week, was too compressed. The first session carried too much theory, which left too little time for choosing a project. The first exercise had a bug: assistants surveyed participants instead of reading their own history. And recording depended on manual commands, so only two of the four plenaries were captured.

The bridge stayed open, and that is the finding. Individual tooling is getting easy, and getting several people's setups to work together is where the difficulty is. That is the same question our research asks about agents in organizations.

Next time. Show-and-tell first, then theory, then more show-and-tell. Inventory, then choose a project, then build it, at one session a day over a calmer week. Plan on about three facilitators for ten attendees, because learning AI is closer to an apprenticeship than a course. Grow through an apprenticeship ladder: come to one workshop, then help facilitate the next. The full facilitator retrospective is in the repo.