Something about how prompts are positioned in the Responses ecosystem doesn’t sit right with me.
The docs describe prompts as long-lived, versioned objects you can share across a project. The Responses API lets you reference them by id, version, and variables. Sounds great. But then those same docs tell you to create, version, and manage prompts… in the dashboard. The Assistants migration guide makes this even more explicit — it contrasts API-managed Assistants with prompts that live in the UI, and the migration steps literally tell you to recreate your assistant bundles in the dashboard and stash the prompt ID in source control.
That’s not a missing endpoint. That’s a control-plane gap. You can’t call something an API primitive when its entire lifecycle is UI-first. In practice, this means there’s no clean GitOps story, no CI/CD sync, no drift detection, no way to promote prompts across environments, and no way to reconcile what’s in the dashboard with what’s in your repo. The API lets you *use* prompts — it doesn’t let you *manage* them.
I know I’m not the only one noticing this. There have been community threads going back months asking for the same thing: programmatic prompt management — create, update, list, delete, the works. The ask keeps coming up.
To be clear — I’m not arguing against the Responses API. I think it’s a solid direction. What I’m pushing back on is being asked to treat dashboard-managed objects as production infrastructure.
If prompts are truly first-class, they need first-class lifecycle management: create, list, retrieve, update, version, delete, archive, export, import, and stable identifiers you can promote across environments. Without that, the honest recommendation is: keep prompts in your own systems and treat OpenAI-hosted prompts as a dashboard convenience, not as the backbone of your control plane.
Right now the messaging is contradictory: prompts are the replacement for Assistants, they’re reusable and versioned… but you manage them in the UI. That makes it really hard to commit to Responses as the long-term foundation for agent systems.
A straight answer from OpenAI would go a long way:
1. Is full prompt CRUD/list/export via API on the roadmap?
2. Or is dashboard-managed prompt state the intended long-term model?
Either answer is fine. The ambiguity isn’t.