AI Video Production Workflow: Plan Before You Generate

Start With the Deliverable, Not the Prompt

Most teams open a generation tool within five minutes of a kickoff call. The instinct is understandable and expensive. A model can produce a genuinely beautiful clip that turns out to be useless because it was rendered in the wrong aspect ratio, at the wrong tempo, or with framing that leaves no room for the caption your brand requires.

A deliverable specification is a single page that answers every question an editor, a freelancer, or a future version of you will ask. Write it before generating anything, and every later debate becomes a factual check instead of a matter of taste.

Include, at minimum:

• Runtime and structure. Total seconds, number of shots, and average shot length. A thirty-second teaser typically wants six to nine shots. A sixty-second explainer wants twelve to eighteen. A two-minute brand film wants thirty or more, which changes your review process entirely because no one can hold thirty unfinished shots in their head at once.
• Aspect ratio and destinations. Vertical 9:16 for short-form feeds, 16:9 for web and presentations, 1:1 or 4:5 for placements inside social feeds. Choose once. Re-framing a sequence for a second ratio is not a crop; compositions that breathe in landscape often suffocate in vertical, and subjects placed on a rule-of-thirds line in one format may sit dead center in another.
• Frame rate and motion feel. Cinematic 24 frames per second with a touch of motion blur reads as film. Crisp 30 or 60 frames per second reads as product documentation. Mixing both in a single edit is instantly noticeable even to viewers who could not name what feels wrong.
• Audio plan. Narration only, narration plus a music bed, or fully designed sound. Dialogue changes the shot list because lip-sync shots need tighter framing, shorter durations, and more generation attempts.
• Text and captions. Burned-in subtitles, end cards, lower thirds, disclaimers. Each element consumes screen time and must be planned into shot lengths rather than squeezed in during the edit.
• Brand constraints. Logo placement and clear space, forbidden color pairings, tone of voice, mandatory product angles, and any claim that legal must approve.
• Delivery formats. A master file, platform-specific exports, a music-free version for localization, and often a silent loop for trade show screens.
• Approval milestones. Who signs off on the script, the style frames, the rough cut, and the master. Vague approvals create the most expensive kind of round trip: full-sequence revisions driven by a comment that arrived after picture lock.

Two hours writing this page routinely saves two days of rework. Generative tools make iteration cheap at the shot level and expensive at the sequence level. Fixing a bad shot is a ten-minute job. Rebuilding a sequence around a new aspect ratio or a new runtime is a rebuild.

The Spooling Principle: Foundation Work Beats Model Choice

There is an old craft routine that maps surprisingly well onto AI production: spooling fresh line onto a reel. It is tedious, nobody films it, and if you get it wrong everything downstream suffers. The analogy is worth unpacking because it separates three things creators constantly confuse.

Line weight is model capability

A heavy line handles big fish but casts poorly. A light line casts beautifully and snaps under load. Models behave the same way. Some are tuned for photoreal motion and longer takes; others excel at stylized animation and rapid iteration; others are strongest at short, tightly controlled camera moves. Choosing a heavyweight pipeline for a quick social cut wastes time. Choosing a lightweight one for a hero shot produces mush.

Match the tool to the tension you expect. A shot that must hold a product label legible for three seconds is under heavy load. A shot of drifting clouds behind a title card is under almost none.

Drag is your iteration discipline

A drag setting applies a controlled amount of resistance. Without it, the line tangles the moment a fish runs. In production terms, drag is your review cadence: how many variants you generate per shot, how quickly you decide, and when you stop. Teams with no drag regenerate endlessly and never reach the edit, which is where a film actually gets made.

Set the limit in advance. Three serious attempts per shot at full specification, then change the approach. That rule alone prevents most runaway projects.

Backing is the reference layer you load first

Experienced anglers spool a layer of inexpensive backing before the expensive line. The equivalent is reference material: a locked style frame, a character description, a palette, a scratch voice track, a product photography pack. It costs very little and it protects the parts of the project you cannot easily redo.

The knot is your prompt structure

The knot is the join between intent and tool. A weak knot fails at exactly the moment of maximum load. A prompt that describes a mood but never specifies camera behavior is a weak knot; it will hold for a still image and slip the moment motion is introduced.

Building a Shot List That Survives Generation

The shot list is your storyboard in text form. Each row describes one generated clip with enough specificity that someone else could generate it without asking you a question. Useful columns:

• Shot number and duration
• Camera: shot size plus movement (medium close-up, slow dolly in, static tripod, handheld follow)
• Subject and observable action, described with one primary verb phrase
• Environment and time of day
• Emotional beat or narrative function
• Model class and risk level
• Status: blocked, motion-approved, detailed, final

Writing prompt blocks that hold across a sequence

A reliable prompt has five slots. Keep them in the same order every time so you can diff two prompts quickly.

1. Subject. Two or three fixed attributes that never change. "A woman in a charcoal wool coat with a canvas tote." Not "a stylish woman," which invites the model to reinvent her every take.
2. Action. One observable verb phrase. Models handle one primary action well and two poorly; they handle three by blending them into something you did not ask for.
3. Camera. Shot size plus movement. This is the slot most creators skip, and it is the slot that most improves consistency.
4. Lighting and environment. Time of day, weather, practical light sources, backdrop texture. Consistent lighting vocabulary keeps shots cuttable.
5. Style. Film stock, palette, lens character, animation style, reference era. Keep this block word-for-word identical across every shot that belongs to the same world.

That last point is the cheapest consistency trick available. Rewriting the style block slightly for each shot is how palettes drift and how a sequence starts to feel assembled rather than directed.

A weak prompt: "A cozy morning scene in a modern kitchen, beautiful light, cinematic."

A working prompt: "Medium shot, static tripod; a matte ceramic bottle on a marble counter; soft window light from camera left with visible gradient falloff; shallow depth of field, 50mm lens character; muted warm palette, fine film grain; no text, no hands."

The second version is boring to read and easy to reproduce. Reproducibility is the entire goal.

Negative constraints and artifact guardrails

Negative constraints matter more in video than in stills because artifacts compound over time. A slightly warped hand in a photograph is an annoyance; the same hand flickering across ninety frames destroys the shot. Build a standing list and keep it in the prompt bible so the whole team reuses it: no warped hands, no flickering text, no duplicated limbs, no sudden jump cuts, no morphing faces, no unreadable logos, no camera shake unless requested.

Add project-specific entries as you discover them. When a shot fails for a new reason, that reason graduates into the standing list.

Matching Shot Types to Model Classes

Think in classes rather than product names, because the tool landscape changes faster than any list of names stays accurate. Five classes cover most work.

Pure text-to-video. Best for establishing shots, weather, abstract backgrounds, particle motion, and environments where exact composition does not matter. Fast, forgiving, and ideal for the blocking pass.

Image-to-video. Best whenever composition matters: product hero shots, character close-ups, anything that must match a storyboard frame or a brand guideline. A practical rule: if you can draw the frame, generate it as an image first, refine it until it is correct, then animate it. You gain enormous control and eliminate most reshoots.

Video-to-video restyling. Best for turning captured footage into an animated look, or for raising the finish of a low-cost plate. Excellent for lifestyle inserts where real hands and real product handling are non-negotiable.

Motion graphics and 2D animation. Best for data, diagrams, typography, and anything requiring precise brand control. Generated animation is wonderful for texture and terrible for legible numbers. Do not ask it to render your pricing table.

Hybrid compositing. Best for the shots that carry the most risk: product interactions, faces, hands, logos. Generate a background plate, generate or photograph the hero element, then composite. It is more work per shot and dramatically less work per project.

Decision criteria for each shot

Ask five questions. How complex is the motion? Does subject identity need to survive a cut? How long is the shot? How much camera control is required? How expensive is a bad take?

High control plus high identity equals image-to-video or hybrid. High velocity plus low identity equals text-to-video. Legible text and exact numbers equal motion graphics, always.

The last twenty percent comes from specialty tools

Finishing quality rarely comes from the generator. It comes from upscaling, frame interpolation for smooth motion, background removal and rotoscoping for compositing, voice synthesis and audio cleanup for consistent narration, and frame repair tools that fix a single glitched frame instead of forcing a full regeneration.

Budget for these in the plan. A project with no cleanup stage will show its seams in the master.

Organizing Assets, Versions, and Handoffs

This is the backing layer in practice. It is dull and it protects everything else.

• Folder structure. Numbered folders per phase: 01-references, 02-raw-generations, 03-selects, 04-audio, 05-composites, 06-masters, 07-exports.
• Naming convention. shot03_take2_v3_vertical. Include shot number, take, version, and orientation. Nobody should ever open the wrong file during an edit because two files share a name.
• Reference pack. Style frames, character descriptions, palette swatches, product photography, brand assets, and the negative-constraint list. Store it beside the project, not inside a chat thread where it will be lost in a week.
• Scratch audio bed. Even a rough voice track with temp music changes how you judge pacing. Lay it down before final generation, not after.
• Transcripts and caption files. Generate them early so caption timings are designed into the edit rather than crammed on top of it.
• Rejected takes. Keep them. A shot you discarded in week one is often the perfect insert in week three, and regenerating it costs more than storing it.

A twenty-minute investment in naming and folders typically saves several hours during the edit, especially the moment a second person joins the project. Handoffs are where sloppy organization becomes visible.

Also decide your storage rule before you start. Generations accumulate quickly, and a project with four hundred raw clips needs a delete policy: keep selects, keep one alternate per final shot, archive everything else to cold storage.

Iteration Passes: From Blocking to Finishing

Random retries feel productive and are not. Structured passes produce a finished film.

1. Blocking pass. Generate low-effort versions of every shot. Goal: rhythm, coverage, and runtime accuracy. Do not chase beauty here. Exit criteria: the sequence plays at the correct length and every story beat has a shot.
2. Motion pass. Rerun shots where camera movement or subject action is wrong. Fix movement before detail, because a beautiful shot with the wrong move has to be thrown away.
3. Detail pass. Regenerate selects at full quality, upscale, interpolate, repair glitch frames. Exit criteria: every shot is technically clean at playback speed.
4. Audio pass. Lock narration, add music, place sound effects at cut points. Sound effects are what make generated motion feel intentional rather than floaty.
5. Finishing pass. Color match across shots, add titles and captions, confirm loudness, export deliverables.

Enforce the three-attempt rule per shot. If three well-specified attempts fail, the problem is the prompt or the model class, not the random seed. Change the approach: switch to image-to-video, composite practically, or reframe so the difficult element is offscreen.

Timebox each pass. A blocking pass that runs three days has stopped being a blocking pass.

Worked Example: A Sixty-Second Animated Product Film

Imagine a matte ceramic bottle, twelve shots, 16:9 master with a vertical cutdown, narration plus music.

Shot	Class	Risk	Notes
1, establishing kitchen	Text-to-video	Low	Fast pass, forgiving
2, bottle hero push-in	Image-to-video	Medium	Composition locked from still
3, condensation macro	Image-to-video	Medium	Still-first, then animate
4, hand lifting bottle	Hybrid	High	Generate plate, composite real hand
5, lifestyle pour	Video-to-video restyle	Medium	Real footage, animated finish
6, label detail	Image-to-video	High	Text legibility critical
7, abstract liquid	Text-to-video	Low	Background only
8, environmental insert	Text-to-video	Low	Weather and light
9, character reaction	Image-to-video	High	Identity consistency
10, product in context	Hybrid	Medium	Plate plus still product
11, diagram overlay	Motion graphics	Low	Never generate
12, end card	Motion graphics	None	Built in editor

Effort concentrates in shots 4, 6, and 9 because they carry identity and legibility. Shots 1, 7, 8, 11, and 12 are fast and forgiving. That distribution is normal. Knowing it in advance prevents the panic that sets in when the first three attempts on shot 4 do not land.

Plan for a vertical cutdown from the start. Shots 2, 6, and 12 are the ones that usually survive a re-frame; wide environmental shots almost never do, so keep the vertical version's subject centered with headroom and footroom available.

Common Mistakes and How to Correct Them

• Generating before specifying. Fix: write the deliverable page first, even if it takes an afternoon.
• Rewriting the style block for every shot. Fix: copy and paste it. Treat it as a locked asset, not a creative sketch.
• Describing moods instead of cameras. Fix: add shot size and movement to every prompt.
• Chasing perfection on background plates. Fix: reserve full-quality passes for identity shots and composite the rest.
• Reviewing clips in isolation. Fix: review in the timeline, with audio, at final speed. A clip that looks weak alone often plays perfectly in sequence.
• Unstructured feedback. Fix: require notes in a fixed format: shot, timecode, problem, suggested change.
• Version chaos. Fix: one shared master folder, numbered versions, and a rule that no file is ever named final-final.
• Ignoring audio until the end. Fix: scratch track in the blocking pass, locked narration before the detail pass.
• Asking generated video to render legible text. Fix: overlay typography in the editor, always.
• No license record. Fix: track every asset, music track, voice asset, and tool used per shot so commercial clearance is provable later.

Review Cadence, Feedback Format, and Collaboration

Run reviews on a schedule rather than continuously. Continuous asynchronous notes fragment the vision and produce edits that satisfy individual comments while weakening the whole piece. One focused review per week, with a written agenda, outperforms daily scattered messages.

A useful review structure: watch the full sequence once without stopping, then go shot by shot. The first pass tells you whether the story works. The second pass tells you what to fix. Mixing the two makes reviewers comment on a shot that will be cut anyway.

Be explicit about roles. Who owns the final call on visuals, on copy, and on legal compliance? Projects stall when three people assume they each hold a veto. Write the names into the deliverable page.

Version freezes matter too. When a sequence is approved, it gets frozen and archived. New ideas go into the next version, not into a live file that three people are editing.

FAQ and Pre-Flight Checklist

How consistent can generated characters be across a sequence? With a locked description block, a still-image base, and image-to-video animation, characters can stay recognizable across a short film. Full-body movement and complex hand interaction remain the hardest cases, so frame shots to minimize them.

Do I need one tool or several? Most finished projects use at least three: an image generator, a video generator, and an editor with audio and caption support. A cleanup tool for upscaling and frame repair is a common fourth.

How long should a generated shot be? Two to six seconds is the practical sweet spot. Longer clips accumulate drift, and shorter cuts assemble easily.

What if the client keeps changing the runtime? Lock runtime before the detail pass. If it changes afterward, rebuild the blocking pass rather than stretching existing shots, which always looks padded.

How do I handle a shot that keeps failing? Change the model class, not the seed. If image-to-video fails, try a composed still with animated overlays.

Should I generate the music too? Only as a temp track. Licensed or composed music clears more reliably and holds up better in the final mix.

What about commercial usage? Check the terms of every tool and asset in your chain, keep a per-shot record, and confirm music and voice rights before delivery.

Before you generate the final pass, confirm: deliverable spec locked, prompt bible consistent, reference pack complete, negative constraints applied, audio bed in place, naming convention in use, review notes resolved, and the export list matched to platform requirements.

Preparation is not the glamorous part of AI video. It is the part that makes every generated frame look like it belongs to the same film.