intermediate
prompts/v1 dies on 30 November: put your prompts in code
Quick answer:
On 30 November 2026 v1/prompts stops existing as an API object. The date goes in the calendar and gets forgotten; what stays is the move: the prompt text stops living in your account and starts living in a repository file.
What shipped
OpenAI is retiring reusable prompt objects from its API. Prompt creation was de-emphasised from 3 June 2026 and v1/prompts shuts down on 30 November 2026, less than a month from the date of this post. The company's prompt-engineering guide says so, and so does its deprecations page. The same June announcement covered the Evals platform and Agent Builder. Six months separate the two dates.
The migration OpenAI proposes fits in one sentence, and it sits where nobody looks for it: "Store production prompts in your application code instead of creating reusable prompt objects. Code-managed prompts let you use typed inputs, code review, tests, and your normal deployment process to change model behavior."
Put plainly for this week: the prompt text has to live somewhere you control. From 30 November, your OpenAI account stops being that place.
What changes
A prompt stops being an object you edit and becomes a file you deploy.
The reason those objects existed is exactly that: change one sentence and watch the effect without touching the repository. The advantage shows up the same day. The cost arrives late: the change went through no review, has no test, and appears in no history you can read six months from now.
With the prompt in a file, changing a sentence is a diff. A diff gets reviewed, deployed and reverted. The difference is not the tooling. It is that for the first time there is something to revert to.
The second change is what "typed inputs" means in OpenAI's sentence. A prompt's blanks stop being template placeholders nobody validates and become arguments to a function. A missing value or a wrong type fails before you spend a call.
What changes for writing prompts
The prompt gains a place in the repository and loses the habit of the quick edit. The first job is an inventory: a prompt that was edited outside the repository has to be found before it can be moved, and those prompts tend to be scattered across services, notebooks and one cron job somebody inherited. The mechanical part of the migration is turning each of them into a file plus a call that reads it; the part that takes thought is deciding what each one is allowed to promise.
Then three consequences. First, the prompt is written the way the code around it is written: in a named file, close to the function that uses it. A prompts/incident-triage.md beside the incidents module gets found; a prompt pasted into a panel does not.
Second, it needs tests that do not check an exact string. A prompt test checks properties: the output invents no fact that was absent from the context, respects the length limit, proposes no action the prompt forbids. The cases live as fixtures, and adding one is cheaper than rewriting the whole prompt.
Third, the git history becomes the record of behaviour changes. The question "why did this assistant change tone in March?" has an answer, and the answer is a diff.
text
# prompts/incident-triage.md
Classify the incident into one of four categories: access, data, billing or
performance. Return only the category and one sentence of reasoning.
Incident: [text]
Affected system: [system]
If the text does not settle it, return "unclassified" and the exact question
that is missing. Do not propose a fix.
What to stop doing
- Editing prompts in a dashboard. A change made there has been seen by nobody but you, cannot be argued about in a review, and leaves no trace when someone asks why the assistant answers differently.
- Keeping the only copy outside version control. A production prompt whose only copy lives in the account has no history, no review and no tests: it has a deadline you do not set.
- Changing behaviour without deploying. If the prompt lives in code, editing it is a deploy and deserves the same treatment: branch, review, deploy, and roll back if the result gets worse.
- Improvising in the console. Trying a variant by hand is fine. Shipping it because the answer came out well is not. The variant becomes a fixture case or it does not stay.
Sources
Summary
Prompt creation in the API was de-emphasised from 3 June 2026 and prompts/v1 shuts down on 30 November. The substance matters more than the date: a prompt stops being an object you edit and becomes a file you review, test and deploy.
Ready to create better prompts?
Try our free AI prompt generator and join thousands of users creating better prompts.