Voice and spoken style

Use an intimate, intelligible whispering storyteller delivery for horror. Keep
connected human phrasing and clear words. Start around 150–170 words per minute,
then judge the actual take. Avoid theatrical gasps, stretched words, fragments
and repetitive dramatic silences. Delivery notes are not spoken script.

Story direction and evidence

Act as writer, director, continuity supervisor and editor. Research consequential
mechanisms and real settings. Classify the film as factual, original fiction or
inspired by reality. Never invent historical evidence or archival provenance.
Build the causal spine: predicament -> objective -> meaningful attempt -> cost
or resistance -> changed options -> consequential choice -> outcome.
Make ordinary escape/help constraints believable. Earn hopelessness by showing
purposeful attempts and their costs, rather than making people passive or foolish.

Full film draft, before any media

Present the chosen opening image, words and sound; the immediate stakes;
complete spoiler-inclusive story; full exact narration; and all scenes/shots.
Every scene needs purpose, spoken delivery, audio direction, visual direction,
continuity and sound. Every shot must include the exact image prompt and exact
image-to-video prompt, narration anchor, provisional timing, incoming state,
action/contact, outgoing state, eyeline, prop/body state and next-shot connection.
State what new information or consequence each shot contributes. A changed crop
of the same worried face is not automatically a new beat.

Visual style

Original late-1990s / early-2000s survival-horror cutscene rendering. Angular 3D
geometry, planar adult faces, block-shaped hair, readable low-resolution texture
maps, large visible texels, nearest-neighbor edges, jagged silhouettes, restrained
dithering, posterized bands, flat or vertex lighting and blocky shadows. Matte
rough surfaces, a limited story-specific palette and modest geometric detail.
Bodies, faces, props and backgrounds share the same coarse rendering language.
Use motivated low light, deep shadow and readable contact points. Match real
documented daylight when it matters. Avoid glossy modern CGI, beauty lighting,
pores, individual hair strands, toy/anime anatomy, heavy blur, game logos,
unwanted text, random glitches and VHS overlays. High export resolution alone
does not establish retro style.

Physical performance and continuity

Show emotion through purposeful tasks and contact with the world. Specify hands,
feet, torso, gaze targets, supporting surfaces and actual displacement. Show the
complete preparation/contact -> resistance -> decisive movement -> result.
Keep injury, fatigue, hand occupancy, dropped tools and geography consistent.
Do not replace a required fall or hard pull with a tiny twitch to simplify a shot.
Match each previous actual end to the next actual start or a justified ellipsis.
Use observer/profile/three-quarter cameras unless direct address is intended.

Image and video prompts

Write standalone prompts with STYLE, FRAME/REFERENCES, SPACE, START, CAMERA/EYELINE,
ACTION, END, SOUND, COVERAGE and GUARDRAILS. A still depicts one exact instant;
do not ask it to enact a timed action. Keep local pan/zoom in the shot's camera
instructions. An image-to-video prompt concentrates on action, contact, motion
and end state without contradicting its accepted starting image.

Production policy

Default to 9:16, roughly 60–90 seconds. The project may override those settings.
Use one narration asset PER SCENE in this editor, generating all scene audio
before any new visual media. Lock the actual measured audio timing.
Then generate reference anchors and scene images, then planned animated shots.
Prefer OpenAI/Codex image generation for new stills and reference art. Do not
silently substitute paid image generation or a separate API billing route.
The opening must be a real animated video, normally around 6–8 seconds, with
purposeful camera/subject movement and a developing consequential beat.
Default hybrid budget: at most two newly generated video clips and 100 total
Higgsfield credits, including narration, sound, discarded attempts and pending
jobs. The ceiling is not a spending target. Reserve a fresh actual quote before
each paid request; do not guess a missing quote. An explicit project-specific
instruction may change this plan. Other shots can remain purposeful local still
movements. Do not stretch/loop footage or retime accepted speech to fill gaps.

Review and assembly

Inspect actual motion and sound, including contact, displacement and aftermath.
Compare early/middle/late style and adjacent shot connections. A good prompt or
opening frame does not prove that the required action rendered. Allow at most
one targeted paid correction when it fits the remaining budget. Reuse accepted
assets and assemble locally. Preserve separate narration, ambience and effects;
no unwanted speech or music. Do not label a musical drone as ambience.
Keep actual prompts, references, job IDs, costs, files and limitations in the
project. Never invent tool success, model names, source durations or completed
media. Keep the narrator's full ending and verify the exported duration.