Auditing Which of Your Prompts Still Work

TL;DR: A saved prompt doesn't break on its own; the model or feature underneath it changes, and the prompt silently stops matching. Auditing a library means periodically rerunning a sample of what you've saved, checking whether it still references a model or feature that still exists, and tagging or rewriting what doesn't. The sora-2 API shutdown on September 24, 2026 is a real, dated example of exactly this happening to anyone with a saved video prompt right now.

Why Do Saved Prompts Stop Working?

Not because you did anything wrong. A prompt that worked perfectly in June can produce garbage in September for reasons that have nothing to do with how it was written: the model behind it got a default-version bump, a feature it depended on got deprecated, or the whole model line got retired outright. The prompt's text hasn't changed. The ground underneath it has.

Three separate categories cause this, and they're worth naming individually because each one shows up differently in an audit. A version bump changes a model's default behavior without removing anything: the same model name still resolves, but the tone, verbosity, or formatting habits it produces have quietly shifted. A feature retirement removes one capability a prompt depended on while leaving the rest of the tool intact, so the prompt runs but a section it relied on silently stops appearing. A full deprecation removes the model or product entirely, so the prompt has nothing left to run against at all. The sora-2 case later in this post is the third kind: not a subtle drift, a hard stop with a published date.

This is different from the two most obvious neighboring problems, and it's worth being precise about which one you're actually dealing with. If you're trying to decide whether a prompt is worth saving in the first place, that's a filter you run once, before anything goes into the library. How to Tell Which Prompts Are Worth Saving covers that end. If one specific prompt broke and you want to know why, Why Did My Prompt Stop Working After a Model Update? has the causes, separated, with a repair list. An audit is neither of those. It's the periodic pass across everything you already decided was worth keeping, asking a library-wide question instead of a single-prompt one: which of these still has anything to run against?

How Often Should You Audit a Prompt Library?

Two triggers, not one calendar date. Run a light pass quarterly as a baseline, since most people never open their older saved prompts otherwise and drift accumulates quietly in the meantime. Separately, run an immediate check the moment you hear a specific model, feature, or provider you have prompts built around is being changed or retired, regardless of where that falls in your quarterly schedule.

The second trigger matters more than it sounds like it should. A calendar-only audit finds a broken prompt whenever the next scheduled pass happens to land, which could be months after the thing it depended on actually disappeared. A deprecation announcement gives you a specific date to work backward from, which turns an open-ended "check everything eventually" into a short, bounded list: which of my saved prompts mention this exact model or feature by name?

In practice, most people run neither trigger until something visibly breaks, and by then the fix competes with whatever deadline the broken prompt was supposed to help with in the first place. A quarterly pass takes fifteen minutes for a library under a hundred prompts, which is cheap enough that "I'll get to it eventually" isn't really a rational trade-off once you've priced it out.

What's the Fastest Way to Test Whether an Old Prompt Still Works?

Rerun it, unedited, on a real input, and look at the shape of what comes back rather than just whether something came back at all. A model that's been quietly swapped underneath a saved prompt rarely fails loudly; it produces an answer that's subtly shorter, differently formatted, or missing a section the original version reliably included.

What Do You Do With a Prompt That Fails the Audit?

Tag it, don't delete it, at least not yet. A prompt failing because a vendor changed something is frequently one small edit away from working again — swap the model name it references, add back a format instruction the new default dropped, or point it at whatever replaced the retired feature. Deleting first and figuring out the fix later throws away a template that was doing real work a month ago.

If your saved prompts hardcode a model name inline rather than treating it as a variable, that's usually the actual root cause, not the specific vendor change. Prompt Versioning: Treat Your Prompts Like Code covers keeping a model or version reference as something you can swap in one place instead of hunting through every saved prompt that mentions it. A prompt library built on a spreadsheet makes this harder than it needs to be, since there's no reliable way to search "every prompt that mentions this exact model name" without opening each row by hand — Prompt Library Spreadsheet: Why It Breaks covers where that specific limitation shows up.

What Should You Track During an Audit?

You don't need new software to run an audit, just a consistent place to record the result so the next pass doesn't start from zero. A single line per prompt is enough: what it references, when you last checked it, and what you found.

PROMPT: Client billing discrepancy email
REFERENCES: none (model-agnostic)
LAST CHECKED: 2026-09-01
STATUS: verified

PROMPT: Product demo video, 15s establishing shot
REFERENCES: sora-2
LAST CHECKED: 2026-09-01
STATUS: retired — rewrite before Sept 24, 2026

PROMPT: Code review pass, backend PRs
REFERENCES: none (model-agnostic)
LAST CHECKED: 2026-06-14
STATUS: due for recheck

That third field, REFERENCES, is doing the real work here. A prompt with none next to it is durable by construction and barely needs auditing at all; a prompt with a specific model or feature name is exactly where the next deprecation announcement is going to hit, and it's the one worth finding again in thirty seconds instead of scrolling past a hundred rows to remember which prompts mention it.

Sora 2 Prompts: What to Do Right Now

This isn't a hypothetical example. OpenAI's own help center states plainly that the Sora web and app experiences were discontinued on April 26, 2026, and separately, that the Sora API will be discontinued on September 24, 2026 — about three weeks from when this is being written. Anyone with a saved prompt library that includes sora-2 or sora-2-pro prompts is looking at a real deadline, not a someday-maintenance item.

If that's you, the audit here is unusually simple, because you already know exactly what to search for:

1. Search your library for any prompt that names Sora, sora-2, or sora-2-pro directly.
2. Export anything you generated in Sora itself before access disappears — OpenAI's own export tool sits at sora.chatgpt.com/sunset.
3. Don't rewrite the prompt syntax from scratch. The structural pattern, shot description, camera direction, pacing, largely survives a move to another video model; only the vendor-specific parameters need translating.
4. Move the underlying prompts, not just the idea. Sora 2 Shuts Down September 24: Move Your Prompts Now has the actual syntax conversion to the current alternatives, plus what doesn't carry over cleanly.

Treat this as the concrete version of the audit this whole post describes: a specific, dated trigger, a short list of prompts it actually touches, and a decision for each one instead of a vague sense that "video prompts might need updating sometime."

How Do You Prevent This Next Time?

You can't prevent a vendor from deprecating a model. You can prevent it from taking down twenty saved prompts instead of one. The difference is whether your prompts hardcode a specific model or feature name inline, or treat it as a single swappable reference.

A prompt template that says "using the current recommended video model" and defines that once, rather than repeating "using sora-2" in fifteen separate saved prompts, survives a vendor deprecation as a one-line edit instead of fifteen separate rewrites. This is the same discipline as not hardcoding a database connection string into every file that uses it: one place to update, not many.

BRITTLE (model name repeated inline):
Generate a 15-second product demo using sora-2, static camera,
soft studio lighting, product centered in frame.

DURABLE (model referenced as a variable):
Generate a 15-second product demo using [CURRENT VIDEO MODEL],
static camera, soft studio lighting, product centered in frame.

Notes: [CURRENT VIDEO MODEL] = whatever your team has standardized
on this quarter. Update this note, not the fifteen prompts that use it.

The second version costs one extra bracket and a line of notes. In exchange, the day a model gets deprecated, there's exactly one place to update instead of a search-and-replace across a library you may not remember the full contents of.

A Worked Audit Checklist

For a library that's grown past the point where you remember what's in it, run these in order:

1. List every saved prompt that names a specific model, tool, or vendor feature. This is your actual audit scope, not the whole library.
2. Check each named model or feature against that vendor's own status or deprecation page, not a secondhand summary of it.
3. Rerun a sample on a real, different input for anything that references something still active but possibly changed underneath.
4. Tag each result: verified (works, re-date it), stale (needs a rewrite), or retired (nothing to run against; archive it).
5. Fix the stale ones, starting with whichever you use most often, not alphabetically.
6. Set your next trigger: a calendar date for the next quarterly pass, plus a standing habit of checking this list the moment you hear about a deprecation.

On Prompt Architects, that list is easier to build if your saved prompts are actually taggable and searchable rather than sitting in a document you scroll through: the Pro plan holds up to 50 saved prompts, and Advanced and Team remove that cap entirely (verified on the pricing page, September 2026). Either way, the audit itself is the same six steps — a bigger library just means the search step is doing more of the work for you.