All guides

Craft 10 min August 2026

The AI Filmmaking Rulebook 14 Rules for Directing Generative Video

The distilled rules behind every other guide in this section — built from a fully instrumented production where every rule exists because something specific failed.

Japesh Singhal

Craft & production

Download skill

Two things became clear after enough failed generations to actually count them: structural fixes hold and textual fixes loop — every problem ever permanently solved was solved by changing the shape of a generation call, never by editing prompt text harder. And a pack of rules that only prevents mistakes produces safe, dead films — continuity is the floor, not the ceiling. These 14 rules are what's left after both lessons were taken seriously.

Where this came from
Built from a fully instrumented production: dozens of generation round-trips, most rejected, many hours. Every rule below exists because something specific failed and the fix generalized.
Continuity is not persuasion
A film can be perfectly continuous and still fail as marketing, sales, or creator work. Before you start generating, name the asset type (ad, demo, founder clip, testimonial, brand film, or creator-native recommendation), then state the claim, the visible proof, and the intended next action in one line each. If the piece lives on a creator surface, name the exact format too - direct address, walkthrough, comparison, reaction, or anecdote - so the model is not left filling that trust grammar with generic ad defaults. Full structure: AI video for marketing & sales.
01

The hard rules

01

Platform, then story, before pixels

Two facts come before the idea: can the viewer leave (and how fast), and will they hear it. Score the story before spending any generation budget — a story problem found here costs a sentence; found after generation it costs the whole shot.

e.g. “TikTok/Reels = 0s exit + sound-on; YouTube skippable = 5s exit — decide those before the idea

02

Set the genre dials before shot planning

Anything left unset is set by the model's defaults, and the defaults are identical for every genre. That's why unset films all look alike.

e.g. “Comedy ad: comedy dials + one commercial dial at the end — never two full dial sets at once

03

Lock the world as pixels

A location appearing in more than one shot must exist as an image before the second shot is generated. Prose cannot lock a place.

e.g. “Empty location plate first ("empty platform, white tile, fluorescent lights") — then every shot references that image

04

One speaker, one line, one intention per clip

Two speakers in one generation causes line bleed. Two intentions in eight seconds renders neither.

e.g. “One speaker, one ~3–5s line, mouth closed after — two speakers in one clip bleed lines

05

Decide the audio route before generating

Native (the video model voices it — one re-roll per fix), post-dub + lip-sync (generate with the line spoken, replace the voice — iteration costs a fraction of a credit), or audio-driven (the model performs to audio you supply). Choose on cost of iteration, not quality alone.

e.g. “Directed delivery or non-English → post-dub + lip-sync (cents/line), not native (full re-roll)

06

Put every constraint where it belongs

Pixels, then parameter, then call shape, then a verification check, then prose — in that order of preference. Cap prose constraints at 8; whatever stays in prose is what you'll loop on.

e.g. “Prop count → plate + verification gate; spoken language → gate — keep prose under 8 constraints

07

Never name the artefact you're avoiding

Most video models have no negative prompt, so "no second pair of glasses" is a positive cue for glasses. Rewrite every forbid as an observable positive state.

e.g. “"sunglasses resting on top of his head, eyes fully visible" — not "no second pair"

08

A regeneration is a delta, not a rewrite

Change one variable. If the fix can't be expressed as one change, the shot needs splitting into two.

e.g. “sunglasses.placement → resting on head — change that field only; print the diff

09

One beat, one review

Never generate the next shot while one awaits review. Batching produces bundled multi-defect feedback, which forces multi-axis fixes — which is what causes regressions.

e.g. “Approve shot N before generating N+1 — language fix and sunglasses fix are separate deltas

10

Look before you claim

Never say a clip is correct unless you actually looked at it. If you couldn't look, say that instead.

e.g. “If you didn't watch the take, say "unchecked" — never mark KEEP from the prompt alone

11

Two failures on one axis means the tool is wrong

Stop editing the prompt. Switch model or pipeline stage instead — and revisit an earlier tool preference once evidence contradicts it.

e.g. “Language failed twice on this model → switch model or stage; stop rewriting the same prompt

12

When a loop is reported, deliver a check, not a paragraph

Don't add a rule to a process unless you can name the specific check that enforces it.

e.g. “Gate: "prop count matches the plate" — not another paragraph of rules

13

Direction review before continuity review

Run the silent test (picture only) and the muted test (performance only) before checking props and wardrobe.

e.g. “Silent test (story readable?) → muted test (who wants what?) → then props and wardrobe

14

Finish is a fixed pass

Identical every time. Improvised last-minute settings are how films ship with clipped dialogue or the wrong pixel format.

e.g. “Cadence → grade → texture → one encode per delivery spec → QC — same order every film

15

Cost-control gates before spend

Verbatim canon lines (not concept bullets); workwear brand mark baked into character sheets; name glossary for ambiguous labels; mix/EDL-only asks never regenerate KEEP picture; SPEAK ends after two syllable fails — mute and dub.

e.g. “Brief says "phone thought" only → RECOMMENDATION line + approval flag, not a silent "locked" VO

16

Ship gates before platform-complete

Frame-snap the EDL; beat inventory matches last KEEP; voice map present; open-music readable from t=0 when required; cutdowns, burned captions, and per-surface loudness exist before CLOSE. A hero master alone is not enough.

e.g. “30s 16:9 KEEP'd → still need planned 15s/6s and 9:16/4:5 cutdowns + caption pass

02

The pipeline

The default order of operations, start to finish. Skipping a stage doesn't save time — it moves the cost of finding the problem later, where it's more expensive:

The default pipeline

Brief → Platform (exit pressure, sound state, hero length) → Story (shape, four movements, score it) → Platform again (full surface list, hero ratio) → Genre (set the dials) → Dialogue (write it, then cut the first sentence of every line) → Design (character sheets, location plates) → Shots (beat list, coverage) → Frame-lock (compose each shot's opening frame as a still first) → Generate (one beat → review → one-variable delta if needed) → Chain (each next beat edits the previous approved frame, doesn't reinvent it) → Voice (to picture, never picture to voice) → Cut (hard cuts, overlapping audio across every edit, reactions) → Mix → Finish.

Frame-locking is the highest-leverage step in the whole pipeline
Compose a shot's opening frame in an image model first — which typically accepts far more reference images than a video model does — then animate it. The video model can no longer invent a background, duplicate a prop, or change a wardrobe mid-shot, because those are already locked pixels. A still costs a fraction of a video generation and you can actually verify it by eye before spending on motion.
03

Continuity is not persuasion

A generated film can be perfectly continuous and still fail as marketing, sales, or creator work. Continuity tells you whether the world holds together. Persuasion asks a different question: what is the claim, where is the proof, and what belief or action is the viewer supposed to leave with?

Asset typeWhat must be visibleIt dies when
Direct-response adThe problem or solved image fast, plus a clear next stepThe offer arrives after the viewer's exit window
Product demoThe product actually doing the thingMood and motion replace understanding
Founder or expert clipSpecificity, authority, and one clean point of viewIt sounds like polished marketing copy instead of a real operator
Creator-native recommendationA believable use moment and proof inside itThe product feels pasted onto a borrowed creator format
Brand film or website heroOne memorable emotional association, not a pile of claimsIt explains too much and leaves no image behind
The mute test gets stricter on sales work
If the claim is speed, ease, quality, trust, or luxury, the picture needs to imply that before the voiceover finishes explaining it. A persuasive clip should still point toward its claim on mute.
04

Review order

Run the cheapest, most fundamental check first, and don't proceed past a stage that fails:

OrderCheck
1Story — the idea, the shape, whether every beat earns its place
2Direction — watch it silent: is it readable as a story with no sound?
3Performance — watch it muted: can you tell who wants what?
4World — does the room/location actually hold together shot to shot?
5Continuity — identity, props, wardrobe
6Persuasion — can you infer the claim on mute, and is the proof visible on screen?
7Platform-native trust — if this is creator-facing, does it feel like a real person would post it?
8Audio — transcript accuracy, noise floor, mix balance, sync
9Delivery — pacing, grade, encode, end card
05

Anti-patterns to recognize and refuse

  • Adding constraints until the performance has no room left — safety and craft compete for the same budget.
  • Rewriting a whole prompt to fix one thing that's wrong.
  • Naming a defect in a prompt in order to avoid it (this makes it more likely, not less).
  • An emotion adjective with no named owner (in a two-person shot, it lands on both people).
  • Directing the speaker and not the listener.
  • Playing each generated clip end to end instead of cutting into it.
  • Crossfades as a default connector between shots.
  • Re-rolling a whole shot to fix one mispronounced word.
  • Generating a second beat while the first is still unreviewed.
  • A sales claim with no proof shot, only voiceover language trying to carry belief by itself.
  • Creator-native pacing with copy that no actual person would say out loud.
  • Claiming a clip is correct without having actually checked it.
  • Passing a multi-person reference photo as a single-character identity reference.