Three prompt dialects, global locks and blocking maps

ONEIRIC splits into a film branch and an asset-fabrication branch, mixes three prompt dialects, and relies on named global locks, prop-by-prop physics and generated blocking maps used as composition references.

Original project

This is a reverse-engineering of the prompt system behind ONEIRIC (higgsfield.studio), measured across 3,477 unique prompts. That is the project's complete history, with no sampling: 137 folders and 11,544 generations read.

Everything below is measured.

First: two crafts, not one

The project splits cleanly into two branches. Mixing them makes any average misleading.

1,050 of the 1,369 ASSETS prompts are for image models. Nothing is filmed there. Characters, locations, props and animals are built, and later enter the video prompts as references. It is an art-department stage done inside the same project.

Total: 2,326 video prompts and 1,151 image prompts.

Video output format: 21:9 in 95%, 4K in 76%, and the dominant duration is 15s (66%). These are long clips, not 5-second ones.

The skeleton (video prompts)

ONEIRIC does not have one template. It has three dialects living side by side.

Dialect A is the same family as Cully Hill Boys: same block names, same order. Dialect B appears to be this team's own invention.

Dialect A blocks, with presence across the 2,326 video prompts

Blocks exclusive to dialect B

— PROMPT — is 23% of the total length. In dialect B the prompt is narrated, not tabulated: one long paragraph with the cuts, the dialogue and the actor direction mixed in chronological order.

The techniques, axis by axis

1. Character anchoring

The system has two reference channels and uses them for different things:

<<<image_N>>> is the big difference from Cully Hill (9%). Here, in 4 of every 10 video shots an image is attached and used as a framing reference. When it is used, each one is declared:

Image 1 = <<<uuid>>> (the START location — KEEP this exact place and angle for
the opening on the terrace with Rudy; use it so the composition, architecture
and angle are correct). Image 2 = <<<uuid>>> (young mage, the rider).
Image 3 = <<<uuid>>> (griffon mount). Image 4 = <<<uuid>>> (mage council —
for their positions).

A reference's role is trimmed with appearance only, not with USE FOR/IGNORE. USE FOR: / IGNORE:, the Cully Hill signature, appears 0 times in ONEIRIC. Instead the character sheet ends with a short phrase that closes the scope (35% of prompts):

<<<uuid>>>: slim 18-year-old man, short (~170 cm), ginger hair, rimless glasses,
navy-and-cream raglan "Quantum" tee, light-blue jeans; sitting in the beige
armchair on the screen-right, a slice of pizza in his hand; a sharp, wiry, nerdy
skeptic who is secretly a huge dreamer. Character appearance only. Reference.

Look at what that sheet does: physical identity, position in the frame and psychology, in one line. The character description is more than a portrait. It carries the dramatic role ("secretly a huge dreamer"), which the performance later cashes in. And Present in the wide only limits which shot the character exists in.

100% match is still present (51%) but lower than in Cully Hill (69%), and Render EXACTLY ONE disappeared (0% vs 5%).

2. Performance

On this axis ONEIRIC is less explicit than Cully Hill on micro-behavior and more explicit on drama.

Where it pushes harder is in directing intent inside the dialogue. The — PROMPT — block directs the performance the way an actor's director would, not as a list:

he FIRST looks toward Rudy — his eyes going screen-left and DOWNWARD, to where
Rudy lies low on the couch off-screen — holding a small pause, letting the
question hang and taking the floor; then he answers, working through an
emotional telling to win the guys over... "No brainer." lands straight at them
like the verdict is obvious; on "Deep future for sure — like, space future."
he presses the case, eyes moving between the guys, selling each beat and
holding their eyes on the last words like a trump card thrown down.

Three things in that passage are pure technique:

• the look before the line (FIRST looks… holding a small pause). The beat before the line is directed, not just the line.
• intent per line, in quotation marks with an action verb (lands straight at them like the verdict is obvious).
• in his established character: the role is re-declared in every shot so the performance does not drift between clips.

Lip control also runs the opposite way from Cully Hill. Instead of prohibiting lip-sync, the state is asserted: Every mouth closed, the room quiet, inside the shot where nobody speaks. It is the same instruction phrased in the positive.

3. Physics and logic

Here ONEIRIC beats Cully Hill, and it is its strongest contribution.

The PHYSICS block (56%) is not generic. It gives weight object by object, for the action in this shot.

Real weight as Rudy shifts from fully reclined up onto one elbow — the couch
cushions give, the guitar rests with its own weight across him and slides
slightly as he moves, his phone stays in his hand. Bob's soda can has real
liquid weight as he lifts it and sips. Sam's pizza slice is a clean, intact
slice, floppy with a grease...

Every prop gets its own weight note. The couch gives, the guitar slides, the can has liquid weight, the pizza slice is floppy. The model is not asked for "realistic physics". It is told what each mass does.

ONEIRIC's invention: GLOBAL LOCKS with declared priority. A named lock, in capitals, with its scope and its rank:

GLOBAL LOOK LOCK — NO VIGNETTE, EVER: absolutely NO vignette anywhere, in any
frame — corners and edges exactly as bright as the centre.        (199 uses)

GLOBAL AUDIO LOCK — SFX ONLY, FOREVER: SFX ONLY — diegetic sound effects only.
NO music, NO score, NO soundtrack.                                (149 uses)

GLOBAL NO ON-SCREEN TEXT / NO WATERMARK LOCK — HIGHEST PRIORITY, ABSOLUTE

GLOBAL BASILISK WINGS LOCK — TWO WINGS, ALWAYS, FOREVER (highest priority):
each basilisk is a WINGED serpent...                               (64 uses)

The grammar is always the same: GLOBAL <WHAT> LOCK — <RULE>, <TIME SCOPE> (<priority>), with words like EVER, FOREVER, ALWAYS, HIGHEST PRIORITY, ABSOLUTE. The block that groups them is literally called — GLOBAL LOCKS (READ FIRST, ABSOLUTE) —, which tells the model the reading order.

For physical action logic there is a separate pattern, the lock that prohibits the geometric alternative:

VERTICAL / TWO-HAND OPEN LOCK (must open VERTICALLY, not sideways): the king
opens the scroll with BOTH HANDS — ONE hand HIGH holding the TOP edge up near
his face, the OTHER hand LOW holding the BOTTOM edge down — so the parchment is
held VERTICALLY TAUT between his hands, hanging TALL top-to-bottom like an open
banner. It is TALLER than it is wide (portrait). FORBIDDEN: he must NOT hold the
two ends left and right, must NOT pull it apart sideways; the sheet must NEVER
be wider than it is tall.

The structure is: lock name, the correct way (with anatomical detail), FORBIDDEN:, the wrong way. The correct mechanic is described with the body (one hand high near the face, the other low) before the wrong one is prohibited.

Frame coordinates are used as in the other two projects (23%), with a different spelling: x 50%, y 55% instead of Cully Hill's x≈42%, y≈68%. What is missing here is 180-degree shutter (0% vs 34%). Geography also leans on screen-left/screen-right (33%) and on the reference image.

4. Framing and camera position

Cully Hill directs the camera in numbers: field of view in degrees in 81% against ONEIRIC's 11%. ONEIRIC directs in lens character and in the reference image:

Shot on a 35mm ANAMORPHIC lens — natural, non-distorting rendering of Bob —
wide open at the fastest aperture for EXTREMELY shallow depth of field and VERY
HEAVY background blur. Only Bob stays razor-sharp; EVERYTHING else — the packed
enemy crush, the background brawl, the passing battering ram, smoke, and all
the foreground/background — is heavily defocused.

When numbers are used, they are used well, including the camera-to-subject distance, which almost nobody writes:

47° diagonal field of view, standard normal lens character, camera 3 to 4 meters
from subject, natural human-eye perspective, comfortable depth of field, the
cramped kitchen readable around him. No distortion.

The most interesting technique in the project: generated blocking maps. 41 image prompts do not produce a still. They produce a schematic staging diagram, which then enters the video prompt as <<<image_N>>>:

<<<image_1>>> — use the attached image ONLY as the compositional guide: copy its
exact framing, camera angle, and the positions, poses and scale of every person
— but do NOT copy its photographic look: no photo textures, no realistic
lighting, no realistic faces, no colors from the image. The OUTPUT is a flat
schematic:

Flat minimalist vector diagram, a BLOCKING MAP for a film scene — an obviously
schematic, non-photographic illustration on a white graph-paper grid background
with thin grey gridlines. Clean flat geometric shapes, solid color blocks, no
shading, no texture, no realism — the look of a technical staging plan.
Front view...: 2 simplified flat human figures. CENTER OF FRAME: a BLUE figure,
kneeling low, facing the camera, head tilted back, mouth open...

In other words, they draw the camera plan with an image model and hand it to the video model as a composition reference. It is the equivalent of a technical storyboard. It solves what text cannot, which is where each body is, at what scale and from what angle, without carrying over light or texture, because the diagram is flat on purpose.

5. Texture

ONEIRIC's texture is built on optics, not stock. Cully Hill says "ARRI Alexa, vintage spherical primes" and trusts that to bring the texture. ONEIRIC describes the behavior of the glass:

Fully photoreal live-action, 35mm filmic look, widescreen frame with a gentle
horizontal squeeze and soft oval bokeh, clean crisp lens glass with a matte
finish, natural depth of field, organic grain, warm cozy grade, grounded
realism. Window and practical light sources render as soft, round, contained
glows that stay within their own sources.

That last sentence is an anti-artifact line. Models tend to let lights bleed and bloom, and the prompt tells them to stay inside their own source. STYLE appears in only 295 prompts but with 23 variants: a locked look, copied verbatim, as in Cully Hill.

Bokeh is specified by shape, not as a word:

ANAMORPHIC LOCK: strong 2x anamorphic — all bokeh tall vertical OVAL / cat-eye,
horizontally squeezed

And for skin:

Absolute photorealism at the micro-texture level, indistinguishable from real
35mm... real skin and hair, the gorilla's individual fur. Filmic contrast
roll-off and real grain.

The technical block always closes the same way: NO CGI. NON-IP. (371 and 212 uses).

The six moves that define ONEIRIC

1. Manufacturing the references inside the project. 1,050 image prompts build characters, locations and props before anything is filmed. Assets are generated, not found.
2. Drawing the staging. The schematic blocking map serves as the composition reference for the video model: geometry without light or texture.
3. Global locks with declared priority. GLOBAL <X> LOCK — …, FOREVER (highest priority), grouped under READ FIRST, ABSOLUTE.
4. Physics prop by prop. Every object in the scene has its own weight note. There is no general rule of "realistic physics".
5. The character sheet carries the dramatic role, not only the physical description, and it is re-declared in every shot with in his established character.
6. Texture comes from optics. The prompts describe how glass and light behave (oval bokeh, contained glows) instead of naming a camera.

Limitations

• The scrape is complete (all unique prompts from the 137 folders), but the API has no ground truth for success: no favorites, no published markers, no parent_id. "What worked" is inferred from the fact that the last version of each shot is the one they kept.
• The three dialects coexist, and block percentages are over the total of video prompts, so a block exclusive to dialect A has a ceiling of 61%.
• 259 image prompts and 135 video prompts are in Russian and follow no template. They read as short retouch requests (for example, "make the reverse of this shot"). Counts of English conventions treat them as absent, which is correct but worth keeping in mind.