ChatGPT’s new custom instructions let you set it and forget it

OpenAI is finally addressing one of the most persistent annoyances of using ChatGPT: having to re-explain your context every single time you start a new conversation. The company is rolling out ‘custom instructions’ in beta, starting with Plus subscribers today and expanding to all users in the coming weeks. Think of it as a standing set of preferences that the model checks before it responds to anything you ask.

The idea is straightforward. You tell ChatGPT once that you’re a third-grade science teacher, or that you prefer coding in a non-Python language, or that you’re meal-planning for a family of six, and it remembers that across all future conversations. No more repetitive setup. OpenAI says this steerability was shaped by user feedback across 22 countries, and the company is framing it as a way to make the model adapt to you, rather than the other way around.

There’s a plugin angle too. If you’ve got a restaurant reservation plugin hooked up and you’ve told ChatGPT what city you live in, the model can pass that location along automatically. That’s a small but meaningful quality-of-life improvement for anyone trying to stitch together a cross-tool workflow. It’s the kind of friction reduction that power users have been jury-rigging with copy-paste prompts for months.

OpenAI is upfront that the beta won’t be flawless — the model might miss instructions or apply them when they don’t belong. They’ve also wired in safety guardrails, using their Moderation API to block instructions that violate usage policies, and the model can refuse or ignore directives that would produce prohibited content. Users can opt out of having their custom instructions used for model training via data controls, and OpenAI says they’ll strip out personal identifiers before any training happens. The feature is accessible on web and iOS for Plus users who toggle it on in settings, though it’s not yet available in the UK and EU.