Negative prompts are often judged by a prettier final clip: the unwanted object disappeared, so the instruction “worked.” That ignores collateral damage. A removal can change identity, flatten motion, erase context, or disturb audio. A fair Veo 3.1 negative prompt test scores the fault and protected scene separately.

Test one protected-detail video brief in APOB

This article provides a four-scene test charter and blank ledger, not fabricated results. Runway documented its negativePrompt field for supported Veo endpoints; Google describes positive prompt elements and parameters. APOB is a place to try an equivalent brief, not evidence that it exposes Runway’s API field. Sources were checked September 11, 2026.

Name the failure before touching the prompt

Do not begin with “make it better.” Name unwanted elements and the details that must survive. The clip cannot prove success if the test keeps changing what success means.

Unwanted-element taxonomy

Classify the target as one of four types: anatomy defect, unwanted text or logo, background clutter, or rigid-geometry error. Give it a location and observation rule. “Extra finger on the right hand between frames 18 and 31” is testable; “bad hands” is not.

Use one fault per paired run. A long exclusion list makes it impossible to know which term caused an improvement or a new defect. Save broader cleanup for a later test.

Protected-detail contract

List what may not change: face anchors, wardrobe color, product label, camera position, action, duration, lighting direction, and expected audio. The contract should fit on a single card beside the source frame.

When working through the APOB AI Video Generator, record the displayed model, mode, aspect ratio, duration, reference image, and prompt. If a control is unavailable, state that limitation instead of inferring a hidden value.

Accept rule

Set a two-part pass: the unwanted element falls below the stated visibility threshold, and every protected detail remains within its own threshold. A removal score of 5 cannot compensate for an identity score of 1.

Example rule: accept only if no extra digit is visible at normal playback or on four sampled frames, while face, sleeve color, cup label, hand path, and spoken line remain recognizable and in the same place.

Stop rule

Stop after the planned baseline, negative-field run, and one descriptive-constraint retest. Stop sooner if a safety filter blocks the source or the asset rights are unclear. The goal is to learn which control path fits the fault, not to hide a hundred failed attempts behind one survivor.

Lock four scenes into paired runs

Prepare a baseline prompt and one reference frame per scene. Copy the positive prompt exactly into its paired run, then add only the negative instruction through the interface being tested. If no seed is available, label the pair “matched inputs, nondeterministic output.”

Hands and anatomy

Use a simple action such as one person lifting a cup with the right hand. Target one unwanted anatomical fault, not “perfect anatomy.” Protect face, hand count, sleeve, cup, contact point, and motion direction.

Baseline template: “Medium shot of one adult lifting a ceramic cup from the table with the right hand, stable camera, natural room light.” Negative-field example: “extra fingers, duplicated hands.” These are test inputs, not reported results.

Text and logos

Use an owned product with one required label and specify one unrelated mark that must not appear. Protect the required label, package color, silhouette, camera angle, and hand placement. If the required label changes, the run fails even when the unwanted mark disappears.

Runway’s API changelog documents the field and supported endpoint behavior. Recheck it before testing because interfaces can change.

Crowded backgrounds

Stage one clear subject in front of a busy but owned background. Name a single removable class—such as stray signage—while protecting depth, room layout, subject outline, and camera motion. Do not ask the model to remove “everything distracting”; that phrase provides no stable observation rule.

This scene reveals whether AI video prompt control removes the named distraction or simplifies the whole environment. Sample the same timecodes in every output.

Rigid geometry

Use a bicycle wheel, table edge, phone, or package with a recognizable axis and corners. Target one common deformation while protecting dimensions, logo position, contact points, and motion. Rigid objects expose collateral changes that a soft background can hide.

Save the first frame, middle action frame, and last frame. A clean first frame does not pass if geometry bends during motion.

Measure removal without rewarding a prettier clip

Review clips without labels such as “negative” or “control.” A reviewer should not know which run is expected to win. Score the target fault first, then inspect the protected contract.

Removal score

Use a five-point scale: 1 means the fault is obvious and persistent; 3 means reduced but still visible; 5 means not visible at normal playback or the sampled frames. Add timecodes. “Looks clean” is not sufficient evidence.

Keep the baseline score beside the tested run. The difference matters more than an isolated number, but it still does not establish causation when seed control is unavailable.

Identity drift

Compare face proportions, hairline, distinctive marks, wardrobe, and product identity. Score drift independently from removal. If identity changes only after the exclusion is added, record the association without claiming the field caused it universally.

For reusable prompt work, keep the evidence in an AI video prompts library with the exact source and model context. A phrase separated from its failed output becomes folklore.

Composition loss

Check crop, subject position, camera height, depth, negative space, and required props. Removal sometimes succeeds by reframing the problem out of view. That is acceptable only if the composition contract allowed it.

Mark “target hidden by crop” separately from “target removed.” The first is a workaround; the second is the behavior being tested.

Motion or audio side effects

Watch once for motion and listen once without looking. Score hand path, camera movement, object contact, tempo, spoken line, ambient sound, and synchronization when audio exists. Google’s official Veo documentation describes prompt ingredients, parameters, limitations, and safety context; use it to define the positive control, not to invent an observed result.

Rewrite the failures as constraints

When the negative instruction removes the target but damages a protected detail—or fails to remove it—rewrite the scene positively. Change one sentence only, then rerun the same evidence protocol.

Descriptive substitute

Replace absence language with the desired visible state. Instead of “no extra fingers,” try “one clearly visible right hand holding the cup by its handle; the left hand remains outside frame.” Instead of “no text,” define “one clean unmarked background sign; preserve the package label exactly.”

Google’s Veo guide emphasizes descriptive prompt elements such as subject, action, style, camera, composition, and audio. That makes it a suitable positive-control reference, not proof that positive wording always performs better.

One-change retest

Copy the baseline and change only the target sentence. Reuse the source image, duration, aspect ratio, and other displayed settings. If seed is unavailable, preserve the limitation in the ledger.

Do not add polish terms during a failure retest. “Cinematic,” “beautiful,” or “dynamic” can change the whole clip and obscure whether the new constraint addressed the fault.

Constraint collision

Look for instructions that cannot all be satisfied: keep a wide view but hide the busy room; preserve the logo but remove all text; show both hands while keeping one outside frame. Rewrite the contract before blaming the generator.

Record which instruction was softened and why. A collision is a brief defect, not model evidence.

Rejected example

Keep at least one rejected clip and annotate the failure at exact frames. A useful rejected-example note might read: “The unwanted wall text is absent, but the required package label also disappears at 00:02.4; reject for protected-text loss.”

That sentence teaches more than a gallery of successful outputs. It also prevents a future editor from reusing the same wording without knowing its tradeoff.

Issue the failure-control card

Turn each observed pattern into a dated decision card. Do not generalize beyond the four scenes, interfaces, and settings you actually tested.

Use negative field

Choose the negative field when the tested endpoint supports it, the target is narrow, the protected contract passes, and the paired evidence shows less collateral damage than the alternatives. Save the exact field value and API or interface version.

This guidance applies to the tested Runway path. It does not imply that APOB exposes a negativePrompt API parameter.

Use positive constraint

Choose a positive constraint when the desired state can be described clearly and its one-change retest preserves more of the scene. Store the sentence with the source and result; do not publish it as a universal magic phrase.

In APOB, use the same protected-detail contract even if the surface presents only a normal prompt. The review method transfers; the underlying control implementation may not.

Change the reference

Replace or repair the source when the unwanted element is already embedded in it, the geometry is ambiguous, the protected details conflict, or both prompt paths fail. A cleaner authorized reference is often more honest than piling exclusions onto a weak frame.

Keep the original case in the archive. Changing the source begins a new test; it does not repair the old pair.

Retest trigger

Repeat the four-scene prompt failure test when the Veo endpoint, documented field, model version, safety behavior, APOB model surface, or acceptance policy changes. Date every card and name its owner.

The final card contains: scene ID, fault, protected details, baseline prompt, negative instruction, descriptive substitute, displayed settings, three scores, rejected frame, chosen control, limitation, and next test date. That is a reusable AI video prompt control asset because evidence travels with the wording.

Sources

• Runway Dev — API Changelog and Veo Negative Prompt

• Google AI for Developers — Generate Videos with Veo 3.1

• APOB AI — AI Video Generator

• APOB AI — AI Video Prompts