By skyforbes I spend a lot of time refining prompts for longer, structured outputs—things like locked formats, step-by-step instructions, and prompts meant to be reused. Over the past couple of weeks, those prompts have felt more stable, with fewer missed instructions or format breaks.
After looking into recent changes, this lines up with the GPT-5.2 rollout.
Here are a few prompt-related behaviors I’ve noticed that might be useful if you care about prompt quality and repeatability:
- Early structure sticks better. When you define sections, steps, or output rules at the top of the prompt, GPT-5.2 does a better job carrying them through to the end.
- Constraints are respected more often. “Must include,” “must avoid,” and formatting rules seem less likely to be ignored halfway through longer responses.
- Simple structure beats clever phrasing. Clear headings, numbered steps, and short instructions work better than dense or overly creative prompts.
- Self-check lines are more effective. Asking the model to confirm it followed all constraints at the end catches more issues than before.
- This isn’t about accuracy. The improvement feels like consistency and follow-through, not fewer factual mistakes. Review still matters.
I didn’t change how I write prompts to see this—it showed up using the same prompt patterns I already rely on.
I wrote up a longer breakdown after testing this across different prompt styles. Sharing only as optional reference—the points above are the main takeaways: https://aigptjournal.com/news-ai/gpt-5-2-update/
For those who build reusable prompts or templates: are you seeing better consistency with longer instructions, or are there still cases where things fall apart late in the response?