October 23 – November 6, 2026
As I've tried to build an always-improving system, one thing I added was a regular sweep to see if we need to make any new skills. This most recent time we didn't find anything new to "skillify" and what we did find was better handled as an addition to an existing process.
Self-improving systems
Back in July I was reflecting on the way the Piper Morgan development process catches its own errors, devises process fixes, and generally exhibits what I called "emergent self-improvement and self-healing behaviors. Some from the model upgrades perhaps but others from our own harness."
I was curious where some of that facility was coming from and we agreed to a monthly skills review to capture things the system is teaching itself and to ratify or "pave the cowpath" where appropriate.
I don't want to accumulate governance for its own sake, so the review was designed to sort ideas into three categories:
- build
- fold into something existing
- don't build
The don't-build option derives from what the agents call my "don't-overlearn principle" which has to do with not overengineering a solution to a rare or low-impact problem that can more easily be handled by existing review and repair processes.
Physician, heal thyself!
Then we forgot to run the review in September or October. (This is ironically the same sort of issue about what it takes for a good idea to actually happen or to be done consistently. Whether or not it requires a skill, it clearly needs a trigger or hook.)
My chief innovation agent (CIO) ensured any future skipped cycle would be visible: the GitHub issue generated by a workflow action as a reminder now gets a comment when a cycle comes due and is skipped.
The manual review
My chief of staff agent (Exec) gathered skill candidates reviewing "Wanted but not found" lines in session logs, the omnibus logs, and its own view across the team from September 1 to October 23. This added ten new candidates to three carried over from July.
Exec offered to discuss with me and I was tired on October 23. They logged the plan to discuss it "tomorrow, when he's fresh" but then I didn't really get to it till November 6 when I reviewed the full set of recommendations and approved them all. Of the thirteen, we folded five into existing processes (two of those were already done by November 4), decided not to build six, and closed two that turned out to be handled already.
So this time nothing new became a skill. At most we needed to add a line to an existing rule, tool, or permission, instead of adding a new procedure.
Even in the time we were evaluating these candidates, two of the recommendations proved to be stale (agents now wake on a schedule, not from session timers, and a push script already existed). CIO caught both by checking against the actual files.
The next time
On December 1 we'll review again. It may turn out that monthly is too frequent. We've already accumulated quite a few skills as it is (some of which have become skills for the Piper Morgan product itself and not just for the operating model and team building it).
I like to think that it's a sign of maturity that our latest round of proposed improvements did not require writing any new skills.
Next on Building Piper Morgan: "The Wall," in which the team runs so hot it hits its weekly usage limit days before the week is over.
When did you last decide not to build something when you realized you didn't really need it? How did you decide?
