Cleaning my agent configurations

After the release of the GPT-5.6 models, and the relaunch of Fable 5, I decided I needed to clean my workflows up. I had countless random skills and plugins lying around from one-off uses. Over time they started creating context noise across projects in Codex. I would ask an agent to do something, read the output, and be unclear why it made a certain choice or did something unprompted. Then I would realize it had loaded an old skill from my AGENTS.md, a plugin folder, or some random markdown document that was still telling it how to work. I wanted to start using less, and let my agents carry the work forward while leaning on what they are already trained to do well. After thinning out my Codex configuration and cleaning up those folders, the results have felt more consistent and easier to understand. There are fewer hidden instructions to compete with the actual task. Another big thing I am steering away from is keeping large markdown files for plans, decisions, and project context. I used to think more documentation would make the work clearer, but it often just creates another layer of abstraction for both me and the agent. The documents get stale, overlap with each other, and quietly become another source of instructions. I would rather keep the important context close to the work, make small durable notes when they are actually useful, and remove the rest. The setup I am moving toward is simple: a small set of global rules, a few skills I use repeatedly, and project instructions that are specific enough to earn their place. Everything else should be temporary or archived. There is also only so much customization you can do inside the Codex app itself when it comes to the harness. That has led me back to trying Pi, where I can go deeper and keep those experiments separate. More on that there. A few metrics I want to track while I do this: - Number of active skills, plugins, and instruction files - How often I have to ask why an agent did something unexpected - First-pass success rate on common tasks - Number of retries or corrections per task - Time spent maintaining configuration versus doing the actual work The goal is not to have the most customized agent setup. It is to have the smallest setup that consistently helps.