← All writing

Writing a better handoff between AI models

I started separating planning from implementation and giving the next model a bounded change with acceptance checks.

I wanted to use a stronger model to plan a change, then hand the implementation to a less expensive one. That sounds straightforward until the second model has to reconstruct the decisions from a long conversation.

On October 11, I asked Sol to plan work that Luna, or another cheaper implementer, could pick up. Sol and Luna were model labels in my setup. The goal was to finish a proof of concept in bounded pieces. Before choosing the next model, we needed to make the work specific enough to hand over.

The plan split the implementation into five slices. Each had file boundaries and checks, with particular attention to persistence and migration behavior. Those details mattered because an apparently small code change could affect what happened to existing saved data. A handoff that only described the desired interface would leave too much for the implementer to infer.

I asked which model to use next. The recommendation was to give Luna one implementation slice, review it, and continue through the remaining slices separately. Sol would handle the final integration. That was a proposed division of work, not a measured comparison showing that both models could deliver the same result at different prices.

What I find useful about this record is the handoff itself. It gave the next session a place to start and a definition of finished. The plan was reported saved before implementation began, so another session wouldn’t need to recover every decision from chat history.

For this kind of delegation, I want the ticket to say what can change and how someone can tell whether it worked. Persistence rules belong there when saved state is involved. The same applies to any constraint that would be expensive to discover after the code is written.

I can’t attach a savings percentage to that decision. The thread doesn’t provide a matched cost comparison or establish how much time I saved. It does record a concrete change in how I assigned work: the model choice came with a defined task, and integration remained a separate responsibility. That gives me something I can inspect when a handoff succeeds or needs another pass.

Written by . Follow via RSS ↗