# The Beat-by-Beat AI Video Production Loop — a prompting skill for AI video production discipline
Source: ekly.ai/guides/ai-video-production-loop — free to use with any AI assistant.

## Hard rules

1. Never go straight from text prompt to moving clip. Lock a still first (composition, framing, positions), review it, then animate.
2. Change one variable per regeneration. If something else also changed, that is a new unreviewed variable — not a successful fix.
3. "One variable" can be a relationship (shared scale, alignment, eyeline). Restage the whole relationship in one generation when one outlier fights the frame's locks.
4. For a position change: state the new position against a fixed landmark in frame, and separately state that the old spot is now empty.
5. Multi-stage gestures need a keyframe chain: lock REST → lock PEAK (middle state) → animate REST→PEAK → animate PEAK→end → concatenate and review the join.
6. When reviewing a new shot, compare back to the first approved shot in the sequence — not only the immediate predecessor.
7. Two failures on the same axis mean change approach (tool or technique), not a third reword of the same prompt.
8. Track the shot list as a dependency graph with parent links (which plate/shot seeded each shot), not a flat numbered list. Walk the parent chain to find drift origins.
9. Once a cut is structurally stable, judge it by the job it has to do — attention, understanding, trust, recall, or click — not by prettiness alone.
10. Dialogue: copy-paste the canonical line from the brief; diff prompt vs brief before submit. Retyping from memory is a full-reroll defect.
11. Concept ≠ script: if only a plan bullet exists ("phone thought"), do not generate it as locked copy — label RECOMMENDATION until approved.
12. N coverage shots listed → N frame-locks (or an explicit drop). Do not collapse a multi-shot beat into one wide silently.
13. Mix / EDL / loudness-only rebuilds: diff beat inventory against the last KEEP'd cut; never regenerate KEEP picture for a mix-only ask.
14. Snap EDL durations to whole frames; cut with frame count, not wall-clock `-t` alone — assert rendered frame count == planned total before muxing audio.
15. Syllable-critical lines (brand word, minimal pair): start on dub, or end the SPEAK route after two wrong pronunciations — do not burn a third native re-roll.

## The loop

LOCATION / COMPOSITION STILL → review → ANIMATE.

Stills are cheaper to iterate; most composition problems are easier to catch before motion.

## Chained-keyframe pattern

Lock REST → Lock PEAK from REST → Animate REST→PEAK → Animate PEAK→REST (or end) → Concatenate, review the join.

Example: hand reaching for a cup — REST (hand at rest) → PEAK (hand gripping cup) → two short clips joined.

## Standing reference

Diff every new shot against beat 1 / the first approved shot. Neighbor-only comparison lets the quality bar drift down invisibly.

## Ledger

Each shot records its actual parent (plate or earlier shot). When drift appears, walk parents backward — the break often sits earlier than the shot where you first noticed it.

## Outcome-fit review

After the loop passes, ask the right question for the clip type:

- Acquisition ad: did proof arrive early enough?
- Product demo: can someone tell what changed and why it matters?
- Founder/expert clip: does the speaker still feel human and credible?
- Creator-native recommendation: would a real person plausibly post this?
- Brand film / hero loop: does one image or feeling stay with the viewer?

## Related guides
- Marketing & sales — ekly.ai/guides/ai-video-for-marketing-and-sales
- Filmmaking rulebook — ekly.ai/guides/ai-filmmaking-rulebook
- Continuity — ekly.ai/guides/ai-video-continuity