There is a particular kind of problem you cannot solve from the outside. You have to be standing inside it while you fix it, like changing a tyre on a moving car, except the car is also the thing you are using to reach the spanner.
That is what building a factory for your own work feels like. I had a loop for turning messy agent runs into something trustworthy. The obvious next move was to use the loop on the loop: point the machinery at itself and see whether it could build the next part of the machinery. It could. That is the good news and the slightly unsettling news, in the same sentence.
You need the factory to build the factory
The bootstrap problem is a small logical knot. To scaffold a new project consistently, you want a workflow. To trust that workflow, you want it to have produced something real. But the first real thing it produces is, awkwardly, part of the factory you are still building.
I could have hand-rolled each new project. I have done that. It produces six subtly different folder layouts, three competing ideas about where decisions live, and a future morning spent wondering why one repo has a session log and another does not. The inconsistency is not a tidiness problem. It is a trust problem. If every project is shaped differently, no agent and no tired human can move between them quickly.
So the first job of the factory was not to build a product. It was to build the thing that builds projects, and then prove it on itself.
What the scaffold actually does
The scaffolding workflow is deliberately boring. Given a new idea, it runs a fixed sequence:
- Classify. Decide what kind of thing this is: a lab, a platform component, a one-off. The classification picks the conventions.
- Create the control plane. A predictable folder layout, a project manifest, and the governance files that make the project inspectable from day one.
- Pre-wire the documentation and the mirror. Session log, decisions, walkthrough, next actions, and a repo-local mirror of the knowledge base, all stubbed and ready.
- Validate. Check the structure against the conventions before a single line of product code exists.
- Stop. This is the step everyone forgets. The scaffold does not start building the product. It hands a clean, documented container to a human and stops.
None of those steps is impressive on its own. Together they mean every project starts from the same shape, with the same trail already laid.
The weekend the labs erupted
The test was not a careful one. Over a single late-May weekend I pointed the workflow at idea after idea, and it produced a row of labs in days: a local model baseline, an architecture map, a publishing lab, the blog factory, a knowledge-base lab, an agent control lab. The factory built its own children, and they all came out the same shape.
The output was not the labs. It was the trail every lab now carried by default.
That is the part that changed how I think about this work. A pile of half-built labs is not impressive; anyone can generate folders. What mattered was that each one arrived already inspectable: classified, documented, validated, with a session log waiting for its first entry. I could open any of them weeks later and orient in under a minute. The scaffold had not saved me typing. It had saved me the much larger cost of not understanding my own work later.
Build the container and the trail, not the product
The lesson generalises past scaffolding. The instinct, when a tool can generate things, is to make it generate the valuable thing: the feature, the article, the deploy. The more useful move is to make it generate the boring substrate that the valuable thing needs to be trustworthy, and then stop.
This is a lab result, not a victory lap. A workflow that produced six labs in a weekend has proven it can create containers. It has not proven those labs will mature, or that the conventions will survive their first real disagreement. Building the factory while standing inside it is a good way to move fast. It is also a good way to pour a foundation you cannot inspect until you are already living on it. The trail is what lets you check.
What Next
Clean Commits in a Multi-Agent World
The scaffold gave every project the same shape. Then several agents started committing to the same repo at once, and the next morning the history no longer added up.