A prompt tuned carefully over months against one model is moved to a newer model from the same family. Output quality drops in ways nobody can characterise precisely. The team wants to prevent this recurring on the next upgrade. What is the most durable response?

The principle — A prompt is written against a model, not in the abstract. Treating it as portable is what turns a routine upgrade into an unexplained regression, and the durable fix is measurement rather than better wording.

Why the key is correct — OpenAI states the problem and the remedy together. On the problem: some prompt engineering techniques work with every model, but different model types might need to be prompted differently to produce the best results, and even different snapshots of models within the same family could produce different results — which is precisely this team's situation. On the remedy, OpenAI strongly recommends two practices for applications of any complexity: pinning production applications to specific model snapshots to ensure consistent behaviour, and building tests and evaluation suites that measure prompt behaviour so you can monitor performance as you iterate, or when you change and upgrade model versions. The two work together. Pinning means upgrades happen when you choose rather than underneath you; the suite means that when you do upgrade, "quality dropped" becomes a set of specific failing cases instead of an impression. Google frames the same discipline more broadly, describing prompt engineering as a test-driven and iterative process.

Why the others are wrong — Writing generically is the strongest alternative and genuinely aids portability, but OpenAI's guidance is that model-specific prompting is what produces the best results, so this buys travel at the cost of the tuning that made the prompt good — and it still offers no way to detect a regression. Freezing the model defers the problem into a larger future version gap with no better instrumentation. And re-tuning then declaring it settled repeats months of manual work while rebuilding exactly the same trap.

Remember this — Pin the version, and make "it got worse" a failing test rather than an opinion.

Sources — OpenAI, Prompt engineering.