How to fix an educational video whose pictures contradict the narration

Last updated 2026-09-22 · 8 min read

What is the short version?

A lesson video is judged by a class of thirty. A pour that misses the glass while the narrator explains pouring is not a cosmetic problem at all; it is the video teaching the opposite of the point, which is why Vizard Agent treats it as a fault rather than a preference.

1. Tell Vizard Agent the timestamps where something is wrong.
2. Say what is wrong at each — the picture, the voice, the match.
3. Expect each fault to be fixed separately, for its own reason.

What do you need before you start?

The video and your list. Vizard Agent will check the whole thing, but a teacher watching with a class spots faults that no automatic check would flag — a character's gender, a demonstration that does not demonstrate, a term used loosely.

• The video. The version you were about to show.
• The timestamps. Where each fault is.
• What is wrong. In your own words.
• What the lesson is. So the fix serves it.
• What must not change. The narration, usually.

What do you type into Vizard Agent?

Describe each fault in one line with its time. Vizard Agent can find where a shot and a sentence disagree, but your reasons — "the speaker is a man, the character is a woman" — are what make the fix unambiguous.

Prompt

Variants worth knowing:

• "It should go into the glass as the narration says." The fault, and the standard.
• "Those should match." Voice and character, stated as a rule.
• "Keep the rest as it is." Scope for a correction pass.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on a school science video that came back from a teacher with three faults listed. The first half of the work is building an index of what the video says and what it shows, before a single frame of it is changed.

1. Downloads and inspects the source video.
2. Builds a visual review sheet of the whole thing.
3. Transcribes the entire audio with timestamps.
4. Maps the shot changes across the video.
5. Reads the transcript and the visual map together to plan the re-cut.
6. Prepares an audio and video synchronisation check.
7. Inspects the exact frames at each reported fault.
8. Listens to the voice consistency across the opening, the middle, the chemistry section and the close.
9. Examines the scenes that need cutting or covering.
10. Assembles all the detail windows into one sheet to review them together.
11. Looks at each problem screen individually — the burning paper, the liquids text, the particle comparison, the flame caption.

Step eight is the step a viewer's complaint rarely names precisely. "The voice changes" can mean the narrator changed, the character changed, or the narration was re-recorded for one section — and the fix differs for each, so the voice is checked section by section.

Step ten is worth asking for by name. Faults reported at three timestamps are usually symptoms of two causes, and seeing the problem frames side by side is what makes that visible.

Step three matters because the narration is the specification. In a teaching video the words are what the lesson is; the pictures are supposed to agree with them, so the transcript is the thing everything else is checked against.

What does the result look like?

The same lesson with the faults gone: the pour lands in the glass, the character matches the voice describing that section, the audio for each section is the right audio, and nothing else in the video has moved.

When does this not work well?

Some faults cannot be patched. If the demonstration is wrong throughout, if the narration itself teaches something incorrect, or if the character design is wrong for the whole video, a correction pass is more expensive than rebuilding the section properly.

• The narration is wrong. The picture is not the problem.
• A fault running throughout. Patch after patch is worse than a rebuild.
• Character design faults. They recur in every scene.
• Generated demonstrations of physics. They will be approximately wrong again.
• No source project. Only the finished file exists to work from.

How do you fix a result that came back wrong?

Keep using the list format you started with. One line, one timestamp, one fault is by far the fastest way to work, because Vizard Agent can act on each line separately without disturbing the corrections it has already made elsewhere.

• "The pour still misses." That shot is replaced rather than retimed.
• "The voice still does not match." The character is changed for that section.
• "You changed the narration." The original audio is restored under the new picture.
• "Now the timing is off." The section is re-synced against the transcript.

How does Vizard Agent compare to doing it yourself?

By hand you fix the timestamps you were given and hand it back, and the class finds two more. The useful move is building the index — what is said, what is shown, where the shots change — and almost nobody does that for a three-minute school video.

Common questions

Can it find faults I did not list? Yes. Ask Vizard Agent to check the whole video against its own narration.

Will the narration change? No, unless you ask — Vizard Agent treats it as fixed.

Can it swap a character's gender for one section? Yes, and Vizard Agent will check the sections either side still match.

What if the demonstration is simply wrong? Vizard Agent replaces that shot rather than adjusting its timing.

Can it re-render only one scene? Yes. That is why giving Vizard Agent timestamps helps so much.

Does it keep the same style? Yes. Vizard Agent matches the treatment of the surrounding scenes.

What about the on-screen text? Vizard Agent checks it against the narration too and corrects any disagreement.

Can it check the science? Vizard Agent flags what looks wrong; the teacher's judgement remains the authority.

Will the video stay the same length? Yes, if you want it to — say so and Vizard Agent works to it.

Can it produce a list of what it changed? Yes, and for a school video that record is worth keeping.

What if two sections use the same shot? Vizard Agent will tell you, because fixing one changes the other.

Does it work from a link? It works best from the file, but a link is enough to start.

Can it add a correction card? Yes, if you are re-publishing something that was already shown.

How fast is a correction pass? Much faster than the original build, since only the listed sections change.