Start With the Edit, Not the Prompt
Most disappointing AI video comes from a workflow that begins in the wrong place. Someone opens a generator, types a lush paragraph about a neon city at dusk, receives a gorgeous eight-second clip, and then discovers it fits nowhere in the video they are actually making. The clip is beautiful and useless. The fix is not a better prompt. The fix is a better sequence of decisions before the prompt.
Before you write a single line of description, answer three questions about the shot you are about to generate.
- What is this shot's job? An establishing shot sells location. A coverage shot carries performance. An insert shot buys you a cut. A transition shot moves the viewer between two worlds. Each job implies a different camera distance, duration, and level of detail.
- Where does it sit in the timeline? A shot that opens a scene can afford a slow push. A shot that lands on a beat needs to arrive already in motion.
- How long must it hold? Four seconds of usable motion is a different problem than twelve. Generators drift over time; the longer the clip, the more likely the face, wardrobe, or background geometry will wander.
Write those answers down. They become constraints, and constraints are what turn a vague prompt into a directed one. A prompt is not a wish; it is a work order. The rest of this guide treats it that way: as a production document that specifies subject, camera, light, motion, and limits, in that order, with as little poetry as the shot can survive.
One more planning habit pays off enormously: decide your aspect ratio and delivery length before generating anything. A vertical short and a widescreen sequence need different framing, different subject scale, and different pacing. Regenerating a finished set of clips in a new aspect ratio is expensive in time and morale. Decide once, early, and build every prompt around that frame.
The Five-Slot Prompt Architecture
Long prompts do not automatically produce better video. Structured prompts do. The most reliable structure that survives across text-to-video, image-to-video, and reference-driven models is a five-slot sentence stack. Each slot answers one production question, and each is short enough to be read by the model without dilution.
Slot 1: Subject and action
One subject, one action, present tense. "A cyclist weaves between stalled taxis" is a shot. "A cyclist, traffic, city life, motion, energy, urban chaos" is a mood board. Models respond to grammar because grammar implies physical relationships. Give them a sentence with a subject and a verb, and you get a subject that does something.
Slot 2: Camera
Specify shot size, angle, lens feel, and movement as one unit. "Medium-wide, low angle, 35mm anamorphic, slow dolly-in" tells the model both where the viewer stands and how the frame changes over time. If you leave this slot empty, you get the default: a locked-off medium shot with mild parallax. Defaults are fine for a first pass and fatal for a finished sequence.
Slot 3: Light and grade
Name the light source, its direction, the contrast level, and the palette. "Overcast dusk, orange firelight rim from screen left, deep shadows, desaturated teal and amber" gives a colorist's instruction. Vague words like "beautiful" or "dramatic" carry no information the model can act on.
Slot 4: Motion and physics
This slot handles speed, weight, and secondary motion: rain, fabric, hair, smoke, dust, embers. Secondary motion is the single strongest cue that a generated clip is real. It is also the first thing lost when a prompt is overcrowded, so give it its own clause.
Slot 5: Constraints and anchors
State what must stay true across the clip: steady tripod framing, subject stays screen right, car remains visible at left edge. Phrase these positively where possible. A positive anchor such as "locked tripod framing" tends to hold better than a negative instruction such as "no camera shake," because the model is being told what the frame is rather than what it is not.
Here is the whole stack in one line:
A woman in a rain-soaked trench coat walks away from a burning car, medium-wide shot, low angle, 35mm anamorphic, slow dolly-in; overcast dusk light with orange firelight rim on her left; heavy rain, coat fabric whipping, embers drifting right to left; locked framing, car stays visible at screen left.
That prompt is about forty-five words. It is not short because brevity is stylish. It is short because every word is doing a job, and nothing is competing for the model's attention.
Why slot order matters
Most models weight the beginning of a prompt more heavily than the end. Put the subject first, camera second, and anchors last. If you find yourself writing three sentences of atmosphere before mentioning who is on screen, you are asking the model to render a mood rather than a shot.
Choosing the Right Model for the Shot Type
Every clip in a finished sequence has a different control requirement. Matching the generation method to the shot is faster than forcing one method to do everything.
Text-to-video
Best for ideation, world building, and shots where the environment matters more than a specific performer. It is the fastest route from an idea to something watchable, and the weakest route for continuity. Use it to explore, and treat its output as look development rather than final footage.
Image-to-video
Best for control. Generate a still with a strong image model, then animate it. Because you approve the composition, wardrobe, and lighting before any motion exists, the failure modes shrink dramatically: you are no longer asking a generator to invent a person and animate them at the same time. For character-driven work, this is the default method.
Video-to-video and motion transfer
Best when performance already exists. Shoot a reference on a phone, restyle it, and keep the choreography. This is how you get believable blocking without asking a generator to invent physical logic from scratch. It is also the most reliable way to hit a specific timing beat, since you control the source performance.
Reference-driven and multi-image generation
Best for series work. When you supply several angles of the same character or prop, the model has more information about identity than a paragraph of adjectives can carry. This approach costs more setup and saves entire reshoot cycles later.
A simple decision rule
Ask what the shot cannot afford to get wrong. If identity matters, use image-to-video or reference-driven generation. If timing matters, use motion transfer. If only atmosphere matters, text-to-video is fine. Then write the prompt in the five-slot structure, because the structure transfers between all of them.
Keeping Characters Consistent Across Shots
Character drift is the most common reason AI sequences fall apart. The face is slightly wider in shot three, the jacket changes color in shot five, and the audience loses the thread. Three practices control it.
Build a character bible
Collect six to ten stills of the same character from different angles and distances: profile, three-quarter, full body, close-up, back. Pick the ones that share identical wardrobe and lighting logic. These become your reference set.
Lock the descriptor text
Write one paragraph describing the character and reuse it verbatim in every prompt. Do not paraphrase between shots. Swapping "silver-framed glasses" for "metal spectacles" sounds like harmless editorial polish, but it gives the model a new token sequence to interpret, and interpretation is exactly where drift begins. Keep a plain text file with locked lines for hair, wardrobe, and signature props.
Use anchors that survive motion
Scars, logos, unusual fabric, a badge, a specific bag. Distinctive features are easier for a model to reproduce than subtle ones. A plain gray jacket will shift shade between shots; a jacket with a visible stitched patch will not.
Patch in post, do not plan around it
If one shot in twenty drifts, a face replacement or relight pass fixes it quickly. If half your shots drift, no amount of patching will rescue the sequence. Fix the cause: reduce the number of variables changing between generations.
Directing Motion: Action, Speed, and Camera Paths
Motion is where prompt writing becomes directing. Three rules cover most of it.
Describe speed with comparators, not adjectives. "She turns her head in roughly half a second" gives the model a duration. "Quickly" gives it nothing. If you want a slow push, describe what happens across the clip: "the frame tightens on her hands over the full shot."
Keep one camera path per clip. A dolly-in and a whip pan are two instructions. Put them in one prompt and the model will compromise, producing a drifting frame that satisfies neither. Choose the move that serves the moment and save the other move for the next shot.
Let physics do the acting. Secondary motion, weight shifting between feet, fabric settling after a turn, dust kicked up on a step. These details sell realism more than facial detail does, and they are cheap to specify in the motion slot.
For high-motion shots, resist the urge to describe chaos with a list of nouns. Instead, describe the trajectory of one object through the frame and let the environment imply the rest. A prompt like "debris crosses from lower right to upper left in front of a stationary camera" reads as motion; "explosion, fire, debris, smoke, panic" reads as noise.
Building Worlds and Styles That Hold Up
Style consistency across a sequence is a documentation problem, not a prompting talent problem. Build a look block, then append it to every shot.
A practical look block names four things: palette, contrast, texture, and lens era. For example: "muted olive and rust palette, medium contrast with lifted blacks, fine 16mm grain, soft vintage glass with visible halation." That sentence does more work than any number of stylistic references, because it describes physical properties that remain stable no matter what the model knows about a particular visual trend.
Avoid stacking styles. Mixing three visual languages in one prompt usually produces a fourth, unplanned one. If you want a hybrid, hybridize through palette and texture, not through competing references.
For larger scenes, add set dressing anchors that repeat across shots: a red awning, a specific bridge, a row of identical lamps. Repeated landmarks are how viewers read geography, and they are how you prove continuity without expensive effects.
A Repeatable Production Workflow, Shot by Shot
Step 1: Beat sheet to shot list
Translate your script into a table with five columns: shot number, job, target duration, prompt slots, and status. Keeping the shot's job visible in the table prevents the common failure of generating beautiful clips that carry no narrative weight.
Step 2: Look development
Generate twenty cheap stills before animating anything. Choose three that define the world. Extract the palette, contrast, and texture language from those three into your look block, then lock it.
Step 3: Prompt drafting and A/B testing
Change one variable at a time. If you alter the camera and the light in the same pass, you learn nothing from the result. Keep a log of what changed and what improved; after fifty shots you will have a personal prompt library worth more than any generic list.
Step 4: Generation passes and selects
Generate four to six variants per shot, then mark in and out points immediately. Selects that sit unfiled for a day become unusable, because you will not remember which take had the clean turn.
Step 5: Assembly, sound, and finishing
Edit before upscaling. Cutting first tells you which frames actually matter, and you avoid spending finishing time on clips that end up on the floor. Add temporary sound early: a music bed and a few impact effects reveal pacing problems that are invisible in silence. Finish with a color pass that unifies all clips, since even consistent prompts will produce slightly different grades.
Common Mistakes and How to Fix Them
- Overloaded prompts. More than six clauses usually reduces control. Fix: cut to the five slots and move the rest into notes.
- Conflicting camera moves. Fix: one move per clip.
- Describing emotion instead of behavior. Models render behavior. Fix: write what the body does.
- Ignoring aspect ratio and safe areas. Fix: decide the frame first, keep key action away from edges where captions or platform UI will land.
- Trusting the first generation. Fix: generate variants and pick deliberately.
- Paraphrasing a locked character description. Fix: copy and paste, always.
- Generating long clips for no reason. Fix: generate short and cut.
- Skipping sound until the end. Fix: add a scratch track early.
A Pre-Export QA Checklist
- Watch every clip at full speed once and at half speed once. Slow playback exposes morphing hands, melting edges, and flickering geometry.
- Check frame edges. Most artifacts hide at the borders, not the center.
- Verify continuity of wardrobe, props, and light direction between adjacent shots.
- Scan for accidental text, logos, and signage. Generators still invent unreadable letterforms.
- Confirm resolution, aspect ratio, and frame rate match delivery specs.
- Confirm the audio bed and any voice track are in sync at the head and tail of each clip.
- Confirm captions do not collide with faces or key action.
FAQ
How many prompts should one shot take? Expect three to six iterations for a hero shot and one or two for simple coverage. If you are on your tenth attempt, the problem is usually in the concept rather than the wording.
Do negative prompts work? They help, but weakly. A positive anchor such as "locked tripod framing" outperforms "no camera shake" almost every time. Rewrite negatives as descriptions of the frame you want.
Why do faces morph over longer clips? Identity is hardest to hold across time. Generate shorter clips, use reference-driven methods, and cut before the drift becomes visible.
Should I generate at a high frame rate? Match your delivery. If the final cut is at 24 frames per second, cinematic motion blur reads better than crisp high-frame-rate motion, which can look like television footage from a different era of production.
Can I reuse one prompt block across different models? Yes for structure, no for results. The five slots transfer, but each model interprets camera language and lighting vocabulary differently. Keep the structure and recalibrate the wording per model.
What is the fastest way to improve? Keep a written log of every prompt and its outcome. Reviewing your own log for twenty minutes teaches more about how a specific model behaves than any general advice, because model behavior changes faster than general advice does.
The through-line is simple: plan the shot, structure the prompt, control one variable at a time, and treat consistency as documentation rather than luck. Do that and your clips stop being impressive demos and start being footage you can cut with.
