Before You Hit Generate: 20-Point Pre-Flight Checklist for AI Video Prompts

Check the decisions before the image exists

I use this checklist when the shot is written and the references are ready, before I commit to generation. It brings together the small decisions that can remain hidden inside a long description: what the sound should do, what an attachment supplies, and where one stretch of action leaves the next.

My working list draws on checks assembled from public Higgsfield material. I treat it as a preparation tool for that workflow, with a yes or no beside every item; if a check fails, I correct the text before generating.

Audio

1. The audio instruction is explicit

• [ ] Audio is disabled, or the environmental-effects instruction below appears exactly as written.

I check the audio setting alongside the words, rather than assuming a visual description settles the sound. For the environmental-effects version of this workflow, I use this line:

Audio: Environmental SFX only. No music. No subtitles.

2. Music is mentioned only in that instruction

• [ ] The word that names music appears nowhere outside the audio line.

Search the complete text, including descriptions of atmosphere. A stray mention can contradict the restricted sound instruction, so I remove it before treating this check as complete.

3. Parentheses have been removed

• [ ] The prompt contains no opening or closing parentheses.

This check comes from the Seedance 2.5 workflow, where the source associates parentheses with a request for composed music. I inspect the actual submitted text, including technical asides, rather than just the main action paragraph.

Scene interpretation

4. An ambiguous scene has a premise

• [ ] If the scene could read as an accident, violence or harm, its opening states what the scene is.

I put the premise before the physical description so the context is present when the action begins. The field is a statement of the intended scene, not a string of denials.

PREMISE:

5. Ages and paragraph boundaries are clear

• [ ] Each person has an age in numbers, and adults and minors occupy separate paragraphs.

This is a check from my video preparation list, not the character-sheet format. I inspect every person described in the shot, so the distinction is not left to an attachment or to the reader's interpretation.

6. The terminology review is complete

• [ ] I have checked the wording for moderation concerns before submitting it.

Mark this only after reviewing the actual text. I keep the wording review distinct from the premise check, because a clear scene description and a reviewed vocabulary are separate parts of preparation.

7. Denials do not introduce unwanted concepts

• [ ] The text contains no defensive denials about children, nudity or resemblance to a real person.

I describe the intended scene directly. The following source examples identify wording to remove; they are not a block to append to a prompt.

not a child
no nudity
not resembling any real person

Names and visible lettering

8. Brand and real-person names are absent

• [ ] The text contains no brands of equipment or objects, film or studio names, or names of directors, cinematographers and other people.

I check the technical description as closely as the action. A name used as shorthand for an appearance is still a name in the submitted text.

9. Lettering has a defined role and language

• [ ] Visible signs use an invented brand or a generic function, and the language of the lettering is declared.

If the shot includes written material, I inspect those instructions separately. The check covers both what the lettering represents and the language in which it should appear.

References

10. The reference set is limited and purposeful

• [ ] There are eight references or fewer, with a different role assigned to each.

Eight is the ceiling used by the source collection behind my list, not a claim about every model's attachment limit. I inspect the role of each reference before deciding that the set is ready.

11. References have written physical descriptions

• [ ] Each reference has its physical description beside it in the prompt.

The attachment and the text must describe the same intended contribution. I read each reference description in place, rather than relying on a separate note that will not be submitted.

12. Identity is separated from sheet presentation

• [ ] Character references include both of the following instructions.

The match instruction concerns the person. The second line stops the studio sheet's presentation from becoming the setting or layout of the generated shot.

100% matches the reference
Studio sheet background and layout NOT inherited

Text preparation

13. The submitted text is entirely in English

• [ ] Action, atmosphere, expressions and reference descriptions are all written in English.

I check the connecting words and labels as well as the paragraphs. A mostly English draft can still contain untranslated directions carried over from working notes.

14. Drafting labels have been resolved

• [ ] There are no unexpanded instruction labels or notes about adding something during editing.

Every instruction that remains must be the instruction intended for generation. I remove preparation shorthand instead of assuming the model will know which missing passage it stands for.

15. Generation settings sit outside the prompt

• [ ] Resolution, duration, aspect ratio and model are in the parameter line, outside the descriptive text.

I compare that line with the intended generation. Keeping the settings together makes this check possible without searching through the action for scattered technical values.

Shot structure

16. The complete header is present

• [ ] The header has been copied intact from the established shot format being used.

I compare the heading with the established shot format, line by line, rather than reconstructing it from memory. Copying it intact keeps this check focused on completeness before I review the individual action stretches.

17. Every stretch starts where the previous one ends

• [ ] Each stretch has a framing and optics heading, and its starting state repeats the preceding final state exactly.

I compare the actual sentences describing those states. A shared intention is not enough if the written positions or object states have changed between them.

18. The visible cast is counted

• [ ] An exact count appears in capitals, together with an instruction excluding unnamed extras.

I reconcile the count with the people described throughout the shot. This check concerns who is visible, so it remains part of the shot structure rather than a general note about the film's cast.

19. Constraints sit beside the action they protect

• [ ] Geography, object state, screen direction and proportion constraints are next to the relevant description, with every negative paired with its positive replacement.

Read each constraint with its neighbouring action. The text should state the required arrangement where it matters, rather than leaving a distant restriction to carry the whole shot.

Generation record

20. The exact submitted text is saved

• [ ] The literal prompt is preserved before generation, ready to record a rejection if one occurs.

I save what will actually be submitted, not an earlier draft. The source workflow notes that a flagged generation may leave no Assets entry, which makes a separate text record important.

Make the decision

I generate only when every applicable check has been resolved in the text or settings. Keeping the exact version lets me investigate a rejection from what I submitted, rather than from what I remember writing.