The project I froze because I designed too much of it
Forge is one of the projects I have complicated feelings about.
I wrote the original specification months before several ideas in it became normal parts of AI coding tools: agents working around a shared project, visual workspaces, browser-like inspection, delegated implementation, review loops, and a layer above individual coding agents that understands the work as a whole.
Then I did what I am very good at doing.
I designed too much of it.
The specification accumulated dependencies on systems that barely existed, some of which I was apparently also supposed to build. The project became increasingly elegant on paper and increasingly difficult to begin in reality.
Eventually I froze it.
The painful part
Watching pieces of that original vision appear elsewhere afterward hurt a little.
Product ideas arriving early has been a recurring pattern in my life. The problem has rarely been imagining the thing.
It has been following through before the world catches up.
I do not want to romanticize that into “I predicted everything.” I did not. Some assumptions were wrong, some parts were overbuilt before they existed, and the tools that eventually appeared solved things in ways I would not have chosen.
But there was enough overlap to make the lesson difficult to ignore.
Being early is not especially valuable if the work remains a document on your hard drive.
What I still care about
The core idea survived.
AI coding tools are becoming increasingly capable individual workers. What interests me is the layer above them: how work is understood, divided, delegated, reviewed, remembered, and presented to the person directing it.
Not another ritual-heavy methodology that makes the human babysit the workflow.
I want the system to make capable agents easier to direct without replacing engineering judgment with process theatre.
The new rule
I am returning to Forge with a much simpler instruction:
Keep the core idea. Remove the empire.
No imaginary dependency graph of products I also need to invent. No requirement that version one solve the future of software engineering.
Find the smallest useful piece that proves the interaction model.
I do not know yet exactly what that becomes.
For once, I am trying to let that be a reason to build something small instead of a reason to design something enormous.