AI Video Generation
Set a quality and verification bar
Quality is a rubric plus independent evidence, not confidence in a polished answer.
Before you start
Why this matters
Without opening an AI tool, write the acceptance test for this job: direct a coherent three-shot paper-boat sequence for a short lesson opener. Name one fact that must be exact, one judgment a person must make, and one condition that should stop the workflow. Compare your answer with the professional standard below; the gap is what you should practice.
1Learn the idea
Read
Set the bar before generation
For direct a coherent three-shot paper-boat sequence for a short lesson opener, define quality across accuracy, completeness, usefulness, safety, and reproducibility. Weight dimensions according to harm. A cosmetic miss can be revised; an unsupported claim, broken calculation, privacy leak, or rights violation blocks release.
Translate each dimension into observable checks. Accuracy means a claim, value, behavior, or frame agrees with an authoritative source. Completeness means every required field or stage appears. Usefulness means viewers watching a clearly labeled generated concept clip can take the intended action. Safety includes the boundary that you must obtain informed consent for any real person's likeness; do not upload private footage, minors, biometric references, or client assets without authorization. Reproducibility means the prompt, input version, settings, and review evidence are saved.
Read
Verification ladder
Use checks from cheapest to strongest:
- Contract check: required sections, schema, length, and prohibited content.
- Source check: trace claims and values to supplied evidence.
- Edge check: run normal, boundary, missing, and adversarial cases.
- Independent check: calculate, test, rehearse, listen, inspect, or open the original.
- Human gate: a responsible reviewer approves consequential use.
In this chapter, the concrete verification is to review first/middle/last frames, check anatomy and physics, compare continuity anchors across shots, inspect logos/text, and verify disclosure and asset licenses. The expected candidate is A bounded shot card with one action, one camera move, stable subject anchors, and negatives that can be checked frame by frame. Record actual evidence, not a checkbox copied from the prompt.
Read
A scoring rubric
Score each criterion 0 (fails), 1 (partly), or 2 (passes). Any zero for factual correctness, permission, privacy, or required disclosure is an automatic stop. A total score is useful for comparing iterations, but it must never average away a blocking defect.
Create a 6-second 16:9 shot: red paper boat in a shallow rain puddle; it drifts past one yellow leaf; low close-up tracking left to right; overcast realistic light; calm continuous motion; no cuts. Preserve boat color and fold shape. Avoid people, text, logos, camera shake, and copyrighted characters.
After generation, sample beyond the happy path. Failures such as too many actions per shot; character drift; warped text; impossible motion; undisclosed synthetic footage; unlicensed music or style imitation often survive a superficial review because the output has the right shape. Use a counterexample designed to expose the riskiest assumption.
Read
Release evidence
Store the rubric result, reviewer, date, input version, failed cases, and unresolved limitations. If the artifact changes, rerun affected checks. Generate short shots separately and edit them together. Consent, copyright, trademark, music rights, and clear labeling are production requirements—not cleanup after export. Quality assurance is part of the work, not an apology added at the end.
Continue learning · glossary & guides
- Which criterion cannot be traded off against a high total score?
- What independent evidence would prove the candidate works in context?
- Reference · Related concept
- Previous
- Next