← All writing

Taking ownership of my AI instructions

Simplifying the operating instructions I wanted to understand and edit myself.

I asked for a review of my AI instructions on November 3 because I wanted to work on them myself. The file had become long and messy, and I wanted better control over what I was asking agents to do. I asked for an analysis before changing it.

The review described routing rules scattered across sections, including instructions that conflicted with one another. Detailed procedures were also sitting in the instructions loaded for every session. Some of that information was only useful during a specific stage of work. Keeping it all together made the operating rules harder for me to follow as well.

I wanted the main workflow to be understandable through a few entry points: plan, build, harden and wrapup. That gave me something concrete to edit. The longer procedures could still exist where they were needed, without every routine becoming another paragraph in the main instruction file.

I also asked for an independent review of my edits using the model labelled Sol 6.1 at medium effort. I wanted another assessment before making the change part of the system I relied on. That was a choice for this review, rather than evidence that the model was the best reviewer in general.

A usage review earlier that day gave this work some context. It identified oversized tasks and incomplete results that required further work from paid coordinating agents. It also found weak completion evidence. Those findings gave me concrete behaviours to address, but the cost data was mostly unknown. I couldn’t turn the review into a credible dollar figure for waste or savings.

My model preference was to use the lightest option that met the quality needed for each task, and to revise that choice as evidence changed. Making that understandable in the instructions seemed more useful than preserving a collection of exceptions added during earlier experiments.

I asked for backups and a staged transition, with clean handoffs before the swap. The thread finished with preparation complete, while the first real cutover and rollback were still untested. That was the honest stopping point for the day. I had reviewed a change to how the system should operate; using it on real work would be the next check.

Written by . Follow via RSS ↗