Prompt Engineering for AI Image and Video Generation

Almost every weak AI image or clip starts with a brief that was never specific enough to fail properly. You type something evocative, the model returns something plausible, and you spend the next twenty minutes swapping words without knowing which word caused which change. The frame is not the problem. The instruction is.

Prompting is the part of generative work that looks like typing and behaves like directing. It is a small skill with a low ceiling on theory and a very high ceiling on execution, and it rewards the same habits a camera operator uses: decide the frame, decide the light, decide what the audience must understand, and remove everything that competes with those three things.

The method below covers a five-block prompt skeleton, a sane approach to negative prompts, a way to keep a character recognizable across ten shots, motion language for clips, and an iteration loop that turns rerolling into something closer to experimenting. Examples work equally well for stills and for video; you can follow along in the AI image generator or the AI video generator.

What a prompt actually does

A prompt is not a search query and it is not a wish. It is a compressed production brief. Every sentence you write either closes a decision or leaves it open, and every decision you leave open gets filled by the statistical average of everything the model has seen. That average has a look: soft rim light, a centered symmetrical subject, shallow depth of field applied indiscriminately, and a palette sitting somewhere between teal and orange. People call it the AI look. It is really the default look of an under-specified brief.

A useful prompt does three jobs at once, and most beginners only do the first:

• Content. What exists in the frame, doing what.
• Camera and optics. Where the viewer stands, what lens they look through, what is in focus.
• Rendering. What medium this appears to be — film stock, print process, illustration style, grain, grade.

If you only write content, the other two are decided for you, and the result will be technically fine and emotionally flat. Compare these:

A detective in a rainy alley, cinematic, 8K, masterpiece.

A tired detective in a soaked wool coat stands under a rusted fire escape, rain falling through a single sodium streetlight above and behind him, 40mm lens at waist height, wet asphalt reflecting the lamp, muted palette of rust and slate, 35mm film grain, night.

The second version is longer, but length is not the point. The point is that every clause in it can be checked against the output. "Cinematic" cannot be checked. "Sodium light above and behind, from the left" can be checked, and if it is wrong you know exactly which phrase to fix.

That is the working definition of a good prompt: a set of instructions specific enough that you can diagnose a failure instead of just feeling disappointed by it.

The five-block skeleton

You do not need a fixed incantation, but you do need a stable order. Blocks make a prompt editable, comparable, and reusable. Sentences blend into each other and become impossible to debug.

The order below also happens to match how most models weight a prompt: early tokens carry more influence, so put the decisions you cannot compromise on first.

Subject and action

Start with the thing that must be true. Not a category — a person, an object, a situation with texture.

"A baker" is a category. "A bakery owner in her fifties, flour to the elbows, pressing dough at five in the morning, sleeves rolled" is a subject. The second version gives the model something to build; the first gives it something to average.

Include only details the frame can show. Eye color matters in a close-up and is wasted in a wide shot. A specific jacket matters if the character recurs. A backstory never matters.

Place, time, and weather

Setting is not decoration; it supplies light, texture, and sound-adjacent cues even in a silent frame. Name the time of day and the weather because both rewrite the lighting for you.

"A bus shelter at dusk in light rain" is a full lighting plan in eight words. "A street" is nothing.

Framing and lens

Choose a distance and a field of view. This is the single highest-leverage change most people never make. Wide lenses exaggerate space and proximity; long lenses compress and isolate; a low camera makes a subject loom; a high camera makes them small.

Say it plainly: "85mm, chest-up, shallow focus" or "24mm, full body, deep focus, camera at knee height." Avoid "professional photography" — it describes an ambition, not a frame.

Light with a direction

Name the source and where it comes from, not the mood. Moods are outputs; sources are inputs.

"Single practical lamp from the left, background unlit" produces a readable, directional image. "Dramatic lighting" produces whatever the model considers dramatic, which is usually a rim light from nowhere. If you want two sources, say so: "window light from the right, dim tungsten fill from a doorway behind."

Palette, texture, finish

Now, and only now, do style. Keep it to one primary anchor and one secondary texture. "Muted palette of rust, slate, and wet black; 35mm grain; slight halation on highlights" is a rendering instruction. "Blade Runner meets Ghibli with Wes Anderson symmetry" is three anchors canceling each other out.

Finish with output constraints: aspect ratio, duration if it is a clip, still versus motion. Deciding the frame shape last is how you end up with a beautiful composition that dies in a vertical crop.

Write nouns, ration adjectives

Adjectives are where prompts go to die. "Beautiful," "stunning," "high quality," "hyper-detailed," and "award-winning" do not describe anything the renderer can act on. They are hopes typed into a text box.

Replace each one with a noun or a measurable property:

• Instead of "beautiful light" → "low sun raking across the floor from camera left"
• Instead of "detailed" → "chipped paint on the doorframe, dust on the sill"
• Instead of "moody" → "two practicals, one burnt out, deep shadow on the far wall"
• Instead of "realistic" → "35mm film, visible grain, natural skin texture, no glow"

A good prompt is mostly nouns doing work and verbs doing more work. "Steam rising off a cup" is worth ten adjectives about atmosphere.

On length: 40 to 90 words handles most single images comfortably. Beyond that you are usually describing two different pictures and should split them. For video, add one sentence for the action and one for the camera move, then stop.

If a phrase matters enough to repeat, it matters enough to move earlier in the prompt. Position is the cheapest emphasis available, and it never breaks syntax.

Negative prompts: name the failure

Negative prompts are subtractive. They remove things. They cannot add composition, fix a flat frame, or rescue a subject that was never described.

The common mistake is dumping abstractions into the negative field — "bad anatomy, ugly, low quality, amateur" — which nudges the model away from a handful of tokens and often flattens contrast as a side effect. The better habit is to list visible, checkable failures you have actually seen in your own outputs.

A practical negative set for a period street scene:

no modern vehicles, no overhead power lines, no visible text or signage in the wrong alphabet, no watermark, no lens flare, no extra people in the background

A practical negative set for a portrait:

no extra fingers, no jewelry, no text, no reflections on glasses, no plastic skin smoothing

Build the list from evidence. Every time you reject an output, ask what specifically was wrong that could have been subtracted, and add that phrase. Within a project you will end up with eight to twelve negatives that are specific to your world, and they will do more than any generic list copied from a forum.

On weighting: if the interface supports it, keep weights mild — roughly 0.8 to 1.3 — and change one at a time. Heavy weights stacked on four tokens produce artifacts. And before you weight anything, ask whether reordering the prompt would achieve the same result more cleanly. It usually would.

Holding a sequence together

One striking frame is easy. Ten frames that read as the same film is the actual job, and it is where casual prompting collapses. Consistency comes from three habits, none of which involve adjectives.

Lock the character paragraph

Write the character once, as a single paragraph, and reuse it word for word across every shot: age range, build, hair, face, one distinguishing feature. Do not paraphrase between generations. Paraphrasing reads as new information, and the model will drift toward it — a different jaw, different hairline, a slightly different age.

Keep wardrobe on a separate line so you can change a coat without rewriting a face. Keep the location in its own short paragraph for the same reason.

Reference stills beat adjectives

A clean reference image does more for consistency than any description. Generate your character in neutral light against a plain background, approve it, and treat it as canon. Do the same for the main location. Then start every shot from the reference rather than from text alone, and use image-to-video when you animate.

When a shot drifts, resist adding more words. Adding description to a drifting prompt usually makes it worse. Re-anchor to the reference and delete the sentence that contradicts it.

Change one block per shot

In a sequence, most prompts should be nearly identical. Vary the framing in shot three, the light direction in shot four, the action in shot five — never all three at once. If something looks wrong, compare the prompt line by line against the previous shot that worked and change exactly one block.

That discipline is what separates a shot list from a folder of nice pictures.

Prompts that move

Video prompts inherit everything above and add a dimension that stills do not have: what changes between the first frame and the last.

A motion prompt should answer four questions, briefly:

• What is the subject doing? One action with a visible start and end.
• How is the camera moving? Slow push in, lateral dolly left, locked-off tripod, handheld follow. Choose one.
• What else is moving? Rain, steam, fabric, hair, traffic, dust in a light beam.
• What is the pace? Deliberate, frantic, drifting, mechanical.

Compare:

A woman in a red coat stands in a snowy street, cinematic.

A woman in a red coat walks toward camera through falling snow, slow push in at chest height, breath visible, streetlights behind her blooming softly, measured pace, night.

The second version gives the model a trajectory. It knows where she starts, where she is going, and that the camera is participating rather than watching.

Two rules keep clips coherent. First, one action and one camera move per generation — a five-second clip fits one beat, and asking for two produces a hybrid pose that reads as a glitch. Second, decide early whether the shot needs a locked first frame. Starting from an approved still gives you far more control over composition, and it makes revisions cheaper because only the motion is in question.

If a sequence needs complexity, build it from simple pieces and cut them together. Structural starting points such as video templates help here, because they force you to think in shots rather than in one long hopeful sentence.

The iteration loop: four variations, one change

The difference between a hobbyist and someone who ships is not talent. It is that one of them rerolls the same prompt ten times, and the other runs four deliberate variations.

Write the shot list first. One line per shot describing what the audience must understand. This prevents the classic failure of generating beautiful frames that do not cut together.

Write the prompt in blocks. Five lines, one per block. Blocks are easy to edit and easy to reuse.

Generate four variations, each changing one block. Variation A changes framing, B changes light direction, C changes palette, D changes texture. Four targeted variations teach you more about a model than twenty random rerolls, because you learn which lever moves which result.

Choose, and write down why. One sentence: "Won because the side light separated her from the background." That sentence becomes a reusable rule. Without it, you will solve the same problem again next month.

Lock the winner as a reference. Everything downstream inherits it.

Animate, then review twice. Watch at full speed for pacing and composition problems, then step through frames for artifacts. Motion problems hide in stills, and composition problems hide in playback.

Mistakes, checks, and a prompt log

Mistakes worth memorizing

Two moments in one clip. "She enters the room and sits down" in five seconds gives you a person half-entering and half-sitting. Split it.

Stacking style anchors. Three aesthetics from three eras produce a frame belonging to none of them.

Using quality words as a substitute for decisions. "8K" never fixed a flat composition.

Choosing aspect ratio at the end. Frame shape should drive composition, not survive it.

Rewriting the whole prompt when one thing is wrong. You destroy the information you already validated.

Ignoring hands, text, and mirrors. These fail most often; write negatives for them before you are surprised.

Keeping nothing. A plain text file of twenty proven prompts outperforms any collection of borrowed keywords.

A two-minute review pass

1. Is the subject readable at thumbnail size? If not, framing is wrong.
2. How many focal points are there? More than one is a composition problem, not a model problem.
3. Does the light have a direction you can name? Flat light means the prompt never named a source.
4. Are hands, text, and reflections clean?
5. Does the palette match the previous shot? If not, re-anchor to the reference.
6. Does the motion read in under two seconds?

Every failed item is a prompt fix, not a taste debate. Reframing review that way makes it faster and considerably less personal.

What belongs in the log

Keep a table with six columns: shot number, prompt version, reference image, output, one-line verdict, and the next change to try. It feels bureaucratic for two shots and saves whole afternoons by the tenth. If you want a sense of how reusable blocks are organized before building your own, browse the prompt library and compare its structure against your notes.

FAQ

How long should a prompt be?

Forty to ninety words covers subject, setting, framing, light, and finish for a still image. For video, add one sentence for action and one for camera movement, then stop. Longer prompts are not better; they are more likely to contain instructions that contradict each other.

Do negative prompts actually work?

Yes, when they name visible failures — extra fingers, watermarks, anachronistic props, reflections on glass. They work poorly as a dumping ground for vague complaints about quality, which mostly dilute the positive description without removing anything.

How do I keep a character consistent across many shots?

Lock the character paragraph verbatim, generate one clean reference still in neutral light, and start every subsequent shot from that reference. Then vary only framing, light, and action. Reusable blocks plus a reference beat any wording trick.

Should stills and video prompts be written differently?

Yes. A still describes a moment; a clip describes a transition between moments. Add explicit motion and camera language, and reduce simultaneous actions so the clip stays coherent.

Why do all my generated images look alike?

Because the prompts are structurally identical. Rotate one variable deliberately — light direction, lens, or palette — and the visual range opens up immediately. New adjectives will not do it.

Is prompting still worth learning as models improve?

Models get better at tolerating loose language, but they cannot invent your intent. The skill is shifting from syntax tricks toward clear direction: shot lists, references, and a consistent visual vocabulary. Those stay useful regardless of which model you open.

What is the fastest way to improve?

Pick one short sequence — five or six shots — and take every shot through the same skeleton. Compare the outputs side by side and write one sentence about each win. Deliberate repetition on a small project beats scattered experiments.

Start directing instead of guessing

Prompting has a short learning curve and a long ceiling. Learn the block structure, ration your adjectives, keep negatives specific, anchor sequences with reference stills, and review against a checklist instead of a feeling. Within a week of deliberate practice, most people stop regenerating the same shot a dozen times and start getting usable frames in two or three passes.

The fastest way to build the habit is a real project. Open the Orelon AI video generator, write a six-shot list, and take each shot through the five blocks — subject, place, framing, light, finish — before you animate anything. Approve your reference stills first, then let motion do the rest. Between sessions, skim the Orelon blog and compare other creators' shot breakdowns against your own log; the gaps are usually one missing block, not a missing model. Cinematic ideas move best when the direction is specific. Write the shot, not the wish.