Dungeon AI Prompts: Build Cinematic Fantasy Worlds

A dungeon is not a room. It is a decision the audience walks into slowly. The corridor, the flooded stairwell, the collapsed crypt — each one is a pressure vessel, and what happens inside it is the story. Most AI-generated fantasy clips miss this completely. They return torch-lit stone and rubble, technically accurate and emotionally inert. The frame looks like a dungeon, but nothing is at stake inside it, so attention slides off in three seconds.

The gap between a clip that feels shot and a clip that feels generated is almost never the model. It is the prompt, written as a shot plan rather than a mood board. A shot plan names a subject, an environment, a dominant light source, a lens, and a turn in the story. Give a video model those five things and it stops decorating and starts directing. This guide walks through a reusable dungeon prompt system you can run inside Orelon or any other text-to-video workflow, from the first line of description to a sequence you would happily cut into a trailer.

Start With Pressure, Not Architecture

Most creators open a dungeon prompt by describing stone. Stone is the least interesting fact about a scene, because stone is constant and people under pressure are not. Before you write a line, answer three questions. What does the character want in this space? What does the space take from them? What is the smallest physical detail that proves it?

A knight losing her nerve grips a sword differently from one who is merely bored. The architecture does not change; the read does. When you build prompts backwards from that pressure — choosing light, framing, and motion to make an inner state visible — clips carry subtext instead of set dressing.

Then translate interior states into observable behaviour. "She is frightened" is not renderable. "She keeps her back to the wall and never looks away from the corridor mouth" is. "He is exhausted" is a label; "he sits on the bottom step with his helmet in his lap, breathing through his mouth" is a shot. That translation habit is the most valuable skill in fantasy prompting, and it is entirely learnable.

One more habit worth building early: write in the present tense. Prompts that describe a moment in progress — "she steps onto the ledge and the rope goes taut" — give a model somewhere to move. Prompts that describe a state — "an ancient dungeon with a warrior inside" — give it a wallpaper to repaint.

Why Light Beats Set Design Every Time

A dungeon is a lighting problem long before it is a set design problem. Viewers forgive an implausible floor plan. They do not forgive a frame where they cannot tell where the light is coming from.

Choose one dominant source

Name its position, its height, its colour, and how fast it falls off. "Warm amber torch at frame left, chest height, heavy falloff into a cool blue corridor" is a decision a model can execute. "Dark and moody" is a wish. Two competing sources of equal strength almost always read as visual noise, so pick a winner and let everything else serve it.

Give the light something to hit

Materials are how light becomes visible. Damp limestone, oxidised bronze, wet flagstones, condensation on iron bars, dust suspended in a beam — each produces different behaviour: soft bounce, specular glints, drifting motes, streaked reflections. A dungeon rendered without any of these reads as a clean, well-ventilated basement.

Keep ambient fill low and motivated

If the ambient fill is generous, the scene looks like a soundstage with a stone texture wrapped around it. Keep fill minimal and justified by something in the world: a distant shaft of daylight, a second lantern just out of frame, the faint blue pulse of a sigil. Motivated darkness is what makes a corridor feel deep.

Decide the palette before the shot list

Palette is a consistency tool disguised as a style choice. Muted teal and amber reads as grounded fantasy. Ash grey and rust reads as survival horror. Cold slate with a single warm accent reads as high-fantasy adventure. Write the palette once and paste the same phrase into every prompt for that location. If the palette drifts, the cuts will feel like they came from different productions even when everything else matches.

A Five-Layer Structure for Dungeon Prompts

Treat every prompt as five stacked layers. You do not need all five in every shot, but you should decide consciously which layer you are skipping — and never skip more than one.

Layer 1: subject, silhouette, and physical state

Silhouette is what the eye registers first in a dark frame, so describe the shape the character makes against the light and the condition they are in. "A hooded figure" is weak. "A hunched figure in a long cloak that drags on wet stone, shoulders forward, right hand gripping a lantern" gives the model something to animate and the editor something to cut around.

Layer 2: architecture with a scale cue

Name the materials, the apparent age, and one structural feature that defines the shot: a collapsed arch, a spiral stair descending out of frame, a flooded floor doubling the torchlight. One defining feature beats six decorative ones because it anchors camera movement. Specify vault height, distance to the vanishing point, or how much of the frame the ceiling occupies, and you control whether the corridor feels endless or claustrophobic.

Layer 3: light, palette, and material behaviour

State the dominant source and what it does. Add one material detail that changes how the beam behaves. Add the palette instruction you locked earlier. This is the layer that stops the render from guessing, and it is the layer beginners skip most often.

Layer 4: camera, lens, and movement

Camera language is the most underused part of fantasy prompting. Specify shot size, lens character, and motion: "slow dolly forward at knee height, 35mm, slight handheld drift." Low angles make dungeons feel enormous. High angles make characters feel hunted. A locked-off wide turns a room into a trap. Even a single camera clause transforms a static render into something an editor can cut.

Layer 5: the emotional turn

Finish with the beat: "the moment before she realises the corridor is looping." That clause changes pacing, expression, and how long the model lingers on a face. It is also the layer that survives a change of tool, which matters when you switch engines mid-project.

Two Worked Prompts, Annotated

Here is the template in plain text. Fill it in, then trim anything that does not change the image.

[Subject + physical state], wearing [specific costume detail], [pose or action]. [Setting], made of [material], with [defining architectural feature]. Lit by [dominant source] with [secondary ambient], [material behaviour detail]. [Shot size], [lens], [camera movement]. Mood: [emotional beat]. Style: [look reference and palette].

A filled crypt example:

A young tomb raider with chalk dust on her sleeves, wearing a leather harness and a lantern hooked to her belt, crouching as she reads a carved inscription. A rectangular burial chamber of damp limestone blocks, with a collapsed ceiling open to a shaft of grey daylight. Lit by the shaft from above as the dominant source, with warm lantern fill from her waist, dust drifting slowly through the beam. Medium wide, 28mm, slow push-in with a slight handheld roll. Mood: the moment she realises the inscription is a warning written in her own handwriting. Style: painterly realism, high dynamic range, muted teal and amber palette.

A flooded-tunnel example:

A tired smuggler in a salt-stained coat, jaw set, wading thigh-deep with a sealed crate balanced on one shoulder. A brick drainage tunnel with a collapsed arch behind him and water the colour of old brass. Lit by a single oil lamp tied to his chest, hard falloff, with a faint green glow leaking from a side grate. Wide, 24mm, slow lateral tracking shot, camera just above the waterline. Mood: the moment he understands the water is rising faster than he can walk. Style: gritty practical-effects fantasy, fine grain, desaturated with one warm accent.

Both prompts are long by chatbot standards and completely normal by film standards. Precision is not verbosity when every clause changes the frame. If a clause would not alter the image, delete it — extra words dilute the ones that matter. Notice also that neither prompt names a monster, yet both feel dangerous. Threat comes from pressure, not from creatures.

Shot Grammar: Four Beats That Make a Sequence Feel Directed

Dungeon scenes have their own grammar. These four patterns cover most of what you need, and they map neatly onto a beat sheet.

The approach

Wide, low camera, character small in frame, architecture dominant. Movement is slow and forward. The point is scale and dread, so keep faces unreadable and let the environment carry the shot.

The threshold

Medium shot at the moment of crossing — a doorway, a ledge, a bridge over a flooded pit. Hold on hands and feet. Threshold shots are where you plant the object the story pays off later: a rope, a key, a broken blade, a wound.

The reveal

Cut to a wide that changes the meaning of the space. The camera moves in one direction only, and the dominant light source should change from the previous shot. If the approach was amber torchlight, make the reveal cold blue or hard white daylight. Contrast does the work, and the audience does the rest.

The turn

The close-up that carries the emotional cost. Keep the camera nearly still and let face and breathing do the work. This is the shot most creators forget to generate, and it is the one that makes a sequence read as a story rather than a location tour.

Four to eight shots is a healthy range for a self-contained beat: an approach, a threshold, a reveal, a turn, plus one or two connective shots. Longer sequences need a second beat sheet, not a longer prompt.

Consistency: Characters, Sets, and the Three-Angle Rule

A single beautiful clip is a demo. A sequence that holds together is a project. Consistency is where fantasy video work usually collapses, and it collapses in three predictable places.

Anchor characters with reference frames

Generate one clean, well-lit reference for each character before you animate anything, then reuse that reference instead of re-describing the person from scratch. Character drift mostly comes from a model reinterpreting adjectives, not from any failure of skill. Locking a reference removes the reinterpretation. You can build those anchors with Create Image and carry them into the video step.

Write a location bible, not a location prompt

Draft one paragraph per dungeon location and treat it as canonical: materials, palette, whether water is present, where light enters, what the floor sounds like underfoot. Every shot prompt for that location then inherits the same five or six facts. Audiences forgive an implausible dungeon. They do not forgive a corridor whose walls change material between cuts.

Apply the three-angle rule

Before calling a set finished, generate the same space from three angles: a wide establishing shot, a low tracking shot, and a close detail. If the room reads as the same room all three times, you have a set. If it reads as three different rooms, solve that before shooting anything else. It is cheap insurance and it saves entire afternoons.

Keep costume language verbatim

Repeating a costume description word for word, in the same order, every time is boring to write and remarkably effective. Paraphrasing is where drift begins. Copy and paste; do not improvise.

Negative Prompts as Tone Control

Negative prompts are usually discussed as artifact cleanup. In fantasy work they matter more as tone control. Worth excluding in a grounded crypt sequence: modern clothing, plastic textures, neon rim lighting, glossy game-engine sheen, symmetric postcard composition, watermark text, and frames stuffed with creatures.

A practical negative line:

modern clothing, plastic, neon lighting, heavy lens flare, text, watermark, extra limbs, blurred faces, oversaturated colour, cartoon shading, cluttered composition, competing light sources

Keep it short. Long negative lists frequently cancel positives you actually wanted. Add one or two exclusions per failed generation, never twenty at once, and log which exclusion fixed which problem. Two weeks of that log is worth more than any prompt pack you can download.

Mistakes, Testing, and How to Judge a Tool

Nine mistakes that flatten dungeon videos

1. Describing a place instead of a moment. Every prompt should contain a present-tense verb.
2. Lighting by adjective. "Moody" tells the model nothing. Name the source, its position, and its colour.
3. Rewriting the character mid-sequence. Write the description once and reuse it word for word.
4. Animating with no camera instruction. The model will invent one, and it will usually be a generic push-in.
5. Ignoring aspect ratio and framing intent. Vertical, square, and widescreen compositions need different prompt priorities.
6. Generating twelve clips before watching one. Watch a single clip, note a single failure, fix a single variable.
7. Treating style as decoration. Style decides consistency more than any other line, so choose it early and lock it.
8. Letting creatures stand in fully lit, clean frames. Partial occlusion reads as menace; full illumination reads as a costume.
9. Cutting on the first frame the model produced. The opening beats of generated motion are soft. Trim into the movement.

Test like an engineer, not an artist

Once prompts become detailed, it gets hard to tell which clause broke a generation. Change one variable per test and keep a four-column log: prompt version, variable changed, result, keep or discard. After ten runs you will have a personal style lock that behaves reliably, and you will stop re-litigating decisions you already made.

The two-engine check is a useful discipline. Run the same prompt through two different generators. If it reads well in both, the prompt is genuinely descriptive. If it only works in one, you have been compensating for that engine's habits with local tricks — fine for one project, risky as a method. Browsing structured examples in the prompt library is a quick way to see how other creators phrase the same ideas before you settle on your own wording.

Criteria that actually matter

Sometimes the prompt is fine and the tool is the constraint. Judge tools on the variables that affect this genre rather than on headline feature lists.

• Motion coherence in low light. Dark frames with a single moving light source are the hardest common case. Test it before committing.
• Identity stability across shots. If the tool offers reference-image or subject-locking workflows, that is worth more than a slightly better single render.
• Prompt length tolerance. Fantasy prompts are long. If a tool silently truncates, your camera instruction may be the part that disappears.
• Iteration speed. You will generate ten to fifteen times more clips than you keep. Speed of the reject cycle matters more than peak quality.
• Aspect ratio options. Widescreen for cinematic sequences, vertical for social cuts — decide which you need before you build a style.
• Export and finishing. Resolution and frame rate should survive a trim, a colour pass, and a sound mix.

A useful habit is to run one test triad at the start of a project: the same shot with a moving light, with a moving character, and with both. Whichever tool handles all three without melting is the one to use for the rest of the sequence.

A Workflow From Beat Sheet to Finished Sequence

1. Write the beat sheet by hand. Four to six lines — approach, threshold, reveal, turn — one sentence each.
2. Lock the style and palette. Pick a look (painterly realism, practical-effects fantasy, graphic-novel contrast) and repeat the same phrase in every prompt.
3. Generate character anchors. One neutral, well-lit reference per character. Iterate here before spending time on motion.
4. Generate location keys. One wide per location to test whether the set reads the way you pictured it.
5. Write shot prompts from the template. One prompt per shot, inheriting the style lock and the location facts.
6. Animate in short batches. Two or three clips at a time. Review, change one variable, repeat.
7. Assemble and cut for rhythm. Trim roughly the first and last half-second of each clip. Let sound design carry the transitions.
8. Bridge unstable shots with stills. If a shot refuses to animate cleanly, animate only the final second of a still image and cut on the movement.

Running that loop inside Create Video keeps iteration fast, and starting from a proven pacing structure rather than a blank page saves time — the templates library is a reasonable place to find shapes you can repurpose for fantasy. If you want to compare how different engines handle the same dark, single-source setup before you commit to one, the alternatives hub is a faster starting point than benchmarking everything yourself.

FAQ

How long should a dungeon prompt be? Long enough to cover subject, setting, dominant light, camera, and mood — usually 40 to 90 words. Shorter than that and you are guessing; much longer and clauses start competing with each other.

Do I need a reference image for every character? For anyone appearing in more than two shots, yes. Text descriptions drift between generations; reference images do not.

Why do my dungeon clips look like game cutscenes? Usually the palette is too saturated and the lighting is too even. Reduce the number of light sources, desaturate, and add one physical imperfection such as dust, condensation, or grit.

Should I write prompts shot by shot or scene by scene? Shot by shot. Scenes are for planning; prompts execute individual shots. Scene-level prompts are the fastest route to a sequence where every clip looks like it came from a different film.

How do I handle creatures and monsters? Keep them partly obscured. Silhouettes, distance, and occlusion read as menace, while a fully lit creature in a clean frame reads as a costume on a person.

Can I reuse one prompt across projects? Reuse the template and the style lock, not the content. Templates scale; identical prompts produce identical clips.

What is the fastest improvement for a beginner? Add a camera instruction and a single dominant light source to every prompt. Those two edits change output more than anything else you can do in an afternoon.

How much sound matters at this stage? More than most creators expect. Dripping water, distant stone shifts, and a low drone will make two mediocre cuts feel intentional. Rough out audio before you judge whether a sequence works.

What if a shot never stabilises? Stop testing it. Replace it with a detail shot — hands, boots, a lantern swinging — and let the audience's imagination bridge the gap. Detail shots are the cheapest way to keep momentum when an engine struggles.

Build the Dungeon, Then Direct It

Dungeon prompts are not a shortcut around storytelling. They are a way of writing the shot list you always wanted to write, in a form a model can execute. Once every prompt contains a subject under pressure, one dominant light source, a camera move, and an emotional turn, the fantasy world stops being a backdrop and becomes a scene.

Start small: one location, four shots — approach, threshold, reveal, turn. Generate the anchors, lock the style, and change one variable at a time until the corridor reads the same way twice. Keep a log, keep your costume language verbatim, and judge tools on how gracefully they handle darkness and identity rather than on how many features they list.

When you are ready to put those ideas in motion, Orelon is built for exactly this: cinematic ideas, shot by shot, until the corridor leads somewhere.