On a construction site, scaffolding goes up before the building does. It holds the workers. It holds the materials. It gives you a surface to stand on while you’re putting the permanent structure together. When the building is done, the scaffolding comes down. If you did it right, nobody thinks about it. The building stands. The scaffolding is gone.

My father ran construction sites. The scaffolding was never the point. The building was the point. But you couldn’t get to the building without the scaffolding. And the quality of the scaffolding determined whether the work above it was safe, stable, possible. Bad scaffolding meant bad work. Dangerous work. Scaffolding that couldn’t hold the load or couldn’t reach the right height meant the crew was fighting the structure instead of using it.

I think about this every time I build a system for someone else to use.

In the classroom, scaffolding was how I got a student from where they were to where they needed to be. A kid who couldn’t process a compound instruction needed the task broken into steps. Step one, then step two, then step three. Each step small enough to hold. The scaffolding was the decomposition. Once the student could do the full task on their own, the scaffolding came off. The goal was always removal. If a student still needed the scaffold six months later, I hadn’t done my job. I’d built a crutch instead of a structure.

FormWork does the same thing for AI tool sessions. It coordinates the session. Names the constraints. Structures the handoffs between tools. When the work is done, the form comes off and the output stands on its own. I named it after concrete formwork for this reason. You build the form, pour the material, wait for it to set, then strip the form. The structure holds because the form held it while it was setting.

The distinction matters because people confuse scaffolding with the finished product. A skill file is scaffolding. A CLAUDE.md is scaffolding. A design system’s documentation is scaffolding. The code it produces is the building. When someone treats the scaffolding as the deliverable, they end up maintaining structure that was supposed to be temporary. The cost goes up. The flexibility goes down. You’re paying to keep the scaffolding standing instead of letting the building do its job.

Good scaffolding is specific to the load it’s carrying. A student with processing delays needs different scaffolding than a student with attention regulation issues. An AI tool processing a complex evaluation needs different scaffolding than one generating a first draft. The scaffold has to match the gap between what the system can do on its own and what it needs to do. Too little and the work collapses. Too much and the system never learns to hold its own weight.

The test is always removal. Take the scaffolding away and see whether the thing stands on its own. If it does, the scaffolding did its job. If it doesn’t, you built a dependency instead of building a capability.