The prompt match score is how the CoreReflex Director knows whether a shot actually rendered what you asked for, not whether it looks nice, but whether it matches the brief. It is one of several concrete checks in the per-shot quality gate, and when a shot misses, the Director reshoots it instead of shipping the miss. Here is how that score is measured, and what it does to your final cut.
What the prompt match score measures
A prompt match score answers a single question for every shot: does the rendered footage contain what the brief asked for? Not "is it pretty" and not "is it sharp". Those are separate checks, but "if the board said a red sports car drifting around a wet corner at dusk, did we get a red sports car, drifting, on a wet road, at dusk?"
The Director boards each shot with three things: a role in the story, a camera move, and a prompt. The prompt match check compares the generated frames back against that prompt and produces a score for adherence. Subjects, actions, setting, time of day, and the key descriptors you wrote all factor in. A shot that quietly swapped the red car for a blue one, or rendered a sunny noon instead of dusk, scores low even when the footage itself is gorgeous.
That distinction, adherence versus aesthetics, is the whole reason the score exists. Generative video models are very good at producing beautiful frames that are subtly wrong. The prompt match score is the guardrail that refuses to let "beautiful but wrong" into your timeline.
Where the prompt match score sits in the quality gate
Prompt match is one check in a per-shot quality gate that scores several concrete things before a shot is allowed into the cut:
- Prompt match, did it render the brief
- Sharpness, is the footage crisp, not soft or smeared
- Motion coherence, does movement hold together frame to frame
- On-screen text legibility, is any text readable and correct
- On-brand, does it respect the brand kit
- Claims-risk, does it avoid unsupported or risky claims
Each check is independent, so a shot can ace sharpness and still fail prompt match. That separation is the point: the gate tells you why a shot is being rejected, not just that it was. If you want the full picture of how the gate fits the larger loop, start with what an agentic AI video director is, and the companion piece on the sharpness gate covers the check that catches soft footage. Both live in our how-it-works library.
How the Director scores a shot
The scoring runs on Vertex AI, the same stack that generated the footage. After a shot renders, the Director samples frames across its duration and uses a multimodal model to compare the prompt's key elements against what is actually on screen. Because the platform owns the full stack, the model judging the shot understands the same prompt that produced it; there is no translation loss between a generator from one vendor and a critic from another.
In practice the check asks a series of grounded sub-questions. Is the named subject present? Is it doing the action described? Is the setting and lighting right? Is anything in the prompt missing, and did anything appear that the prompt never mentioned? Those answers roll up into the adherence score for that shot. The result is concrete and explainable rather than a vague vibe rating, which is exactly what you need when you have to defend a cut to a client or a brand team.
What happens when a shot misses
When a shot fails the prompt match threshold, the Director does not throw away your film and start over, it runs selective regeneration: it regenerates only the failing shot, keeps every shot that passed, and tries again, often with the prompt or parameters nudged to close the gap it detected.
Two design choices make this safe:
- Continuity is preserved. The last frame of the previous shot anchors the next, so a regenerated shot still cuts cleanly into the ones around it instead of jumping.
- Credits stay efficient. Planning, scoring, and editing are free; only generation costs credits. So the gate can be strict without making every reshoot expensive, you only pay to regenerate the shot that actually missed, not the whole sequence.
The payoff is a cut where every shot has cleared the same bar before you ever press play.
Why a score beats eyeballing a generation
Reviewing AI video by eye does not scale, and it is not consistent. A tired reviewer at 6pm approves things a fresh reviewer would reject at 9am. A numeric, per-shot prompt match score gives you the opposite: a consistent, repeatable bar applied to every frame, whether you generate one video or a thousand.
It also changes what review is. Instead of scrubbing a timeline hunting for the off-brief shot, you read the scores and jump straight to the one that needs attention. When you do want to inspect output rigorously, the workflow in how to audit an AI-generated video builds directly on these scores. And if you are weighing platforms, the prompt match gate is one of the capabilities worth comparing in the best AI video generators of 2026, many tools generate, but far fewer score adherence and reshoot the misses for you.
Reading the score in your provenance trace
Every generation carries a portable provenance trace: the model, the prompt, the parameters, and the quality-gate scores, prompt match among them. That trace travels with the asset, so months later you can open a shot and see exactly why it passed and what produced it.
Because the trace is reproducible, you can replay it on the deterministic render path and get the same cut back, frame for frame, the mechanics of which are covered in the deterministic render path. And once a cut is assembled, it lands as a faststart MP4 that plays instantly on the web. The score is not a throwaway number on screen; it is a permanent, auditable part of how the shot was made.
Frequently asked questions
How does the director know a shot matches the prompt?
After a shot renders, the Director samples frames and uses a multimodal Vertex AI model to compare the prompt's key elements, subject, action, setting, lighting, descriptors, against what is on screen, that comparison produces an adherence score. Because the same owned stack generates and judges the footage, the critic understands the exact prompt that created the shot.
What happens if a shot misses the brief?
The shot fails the prompt match check and the Director selectively regenerates just that shot, keeping every shot that passed. Continuity is held by anchoring to the previous shot's last frame, and because only generation costs credits, reshooting one miss is cheap rather than a full restart.
Can I see the prompt match score for each shot?
Yes. The score is part of each shot's provenance trace alongside the model, prompt, and parameters, so you can read it per shot rather than judging the whole film by eye. That makes review fast, you go straight to the shot that scored low.
Does a low score always mean the shot is discarded?
A score below the threshold triggers selective regeneration rather than an outright discard, the Director tries again to hit the brief. The goal is a cut where every shot has cleared the same adherence bar, not a folder of rejected takes.
See the gate in action
The prompt match score is the difference between footage that looks right and footage that is right. Board a film from a single sentence, watch the Director score and reshoot the shots that miss, and inspect the trace behind every frame. Start free with no credit card and run your first scored cut, or read the full mechanics in the developer docs.