I don’t start by asking an agent to build.

I spend real time on the idea first. What the workflow actually looks like for the people inside it, where the friction is, what already exists and why it has not solved the problem, and which assumptions I am carrying that I have not checked. Some of that is research. A lot of it is arguing with the idea.

Not everything survives this. Some ideas turn out to be a feature of something else, or a problem people have already stopped caring about. Working that out early is cheaper than working it out after implementation.

What I want from this stage is a clear view of the problem, the product decisions I already know I have to make, and the specific thing about the idea that is still uncertain.

Once the slice gives me enough confidence to keep going, the work becomes much more deliberate. I define the intended behavior, the product decisions, the scope, the boundaries, and the acceptance criteria. Larger implementations get broken into phases that mean something on their own.

Each phase needs enough specification and context behind it that an agent can work for a long stretch without stopping to ask what I meant. What I am trying to avoid is the point where it runs out of specification or context and starts filling in the blanks itself.

I also turn the ways of working I repeat into reusable structure for the agents, so the method does not have to be rebuilt from scratch on every project.

Long-running work needs the project to still be there hours later, and it needs somewhere for decisions to live that is not a transcript.

That problem is a big part of why I ended up building Fraimed and Sophie in the first place. Without that structure, long agent runs turn into constantly rebuilding the project for the agent, or letting it drift away from the product I intended.