I am interested to know if and how others are using skills in codex. I have been using them to try to streamline repeated prompting and to improve compliance with project development/testing/reporting policies. For this, it works admirably, but at the expense of response speed and task-specific tokens available after skill indoctrination. I still see material degradation of performance after multiple context window compressions, so I am still trying to minimize the project-specific policies and architecture docs needed for a task.
To start, before skills, I had a file, AGENT_PROMPTS.md that I just had opened and copied prompts from depending on the task. This wasn’t sustainable because the easiest to copy and modify prompts were always the most profligate in terms of time and tokens.
Then I tried skills which I considered to be similar to my prompt file, but with codex integration and better controls for dynamically creating the prompt.
And, based on the name, at first, I treated them as skills, for example, “Using the $acme-dev skill on branch b, change the behavior of …” or "Using the “$acme-dev skill refactor …” This seemed to work reasonably well, but started to suffer when I had to augment the skill with a broad range of, well, skills. For instance, how to update backlog, how to report back to me results, how to structure commits. Soon I was back to my omnibus prompt.
Recently, I changed to treating them like roles and this certainly improves the naturalness of prompting them, because I just say, “$acme-dev find the source of the UI flashing when …” - which is not that different, but as i start to develop more roles, I do things like, $acme-pm give me the development report for the last 24 hours and update me as to the state of the backlog." And, probably where I am starting to see real value is when I do things like, “$acme-pm work with $acme-architect to build a multi-agent development plan.”
This has worked better thus far, but the most important aspect of this working was building and modifying the skills as a set. So the skill-creator is always focused on aligning the acme-* skills as a set and never editing one skill without the understanding that the set represents a division of labor.
I would like to know how others are using skills in their projects and whether or not they are making individual skills at all or skills with structured options.
Automations work best with skills and I would use them all the time if I could control model complexity, but it seems like this is not possible.