10 original recipes · Updated 05 Sep 2026

Prompts built for
what happens next.

Realtime video needs more than one beautiful shot. These recipes preserve a world, translate audience intent, expose game state, recover a starving buffer, and keep claims out of generated pixels.

Before you copy

Three prompt systems,
kept separate.

A useful prompt starts with the endpoint contract. Mixing a Director control message with a clip-generation prompt is a product error, not a wording issue.

01

Director configure

A durable world contract sent once at session start: identity, location, camera, audio, continuity, and opening beat.

Continuous WebRTC session
02

Director prompt

A small delta sent during the session. It changes one upcoming action while preserving everything already established.

Mid-stream steering
03

H3 clips

A self-contained, timed shot or a continuation conditioned on a prior frame. It must carry the anchors required for that clip.

T2V · I2V · FL2VA · Ref2VA
01

Native realtime · configure

Director Session Opener

Give a continuous Director session a clear visual contract before the first segment begins.

Model
H3 Max Director
Mode
Configure message
Use when
Starting a new continuous WebRTC session that will receive smaller steering prompts later.
RETIVAI RECIPE / 01
Create a continuous original live-action stream titled [SHOW_TITLE].

WORLD: [ONE-SENTENCE LOCATION], at [TIME OF DAY], with [WEATHER / AMBIENCE]. The world remains physically coherent and time moves forward without resets.

SUBJECT: [CHARACTER NAME], [AGE RANGE AND APPEARANCE], always wearing [WARDROBE ANCHOR] and carrying [PROP ANCHOR]. Preserve face, hair, clothing colors, body proportions, and prop identity throughout the session.

CAMERA CONTRACT: [CAMERA FORMAT AND LENS FEEL]. Favor motivated, continuous movement. Do not jump to a new location unless a later direction explicitly asks for a transition.

OPENING BEAT: [WHAT IS HAPPENING AS THE STREAM BEGINS]. End the opening beat with ongoing motion and room for the next direction.

AUDIO: Natural location ambience plus [VOICE / MUSIC RULE]. Keep speech sparse and only speak words placed inside quotation marks.

CONTINUITY RULES: New directions change only the requested action. Everything else—character identity, wardrobe, location logic, lighting direction, and current objects—persists. No titles, logos, subtitles, watermarks, or copyrighted characters.

Variables

[SHOW_TITLE]
A short internal title that anchors the format.
[ONE-SENTENCE LOCATION]
One concrete place, not a list of possible sets.
[WARDROBE ANCHOR]
Two or three highly visible, persistent clothing details.
[OPENING BEAT]
An action already in motion when playback begins.

Continuity controls

  • Keep persistent facts compact enough to repeat mentally in later directions.
  • Give the character one wardrobe anchor and one prop anchor; too many traits compete.
  • Treat future prompt messages as deltas, not replacements for the world contract.

Audio

Name one ambient bed and one speech rule. Quoted dialogue is more controllable than a general instruction to talk continuously.

Avoid

Multiple locations in the opener, rapid montage language, conflicting camera styles, text on screen, or instructions that depend on an unseen backstory.

Common failure modes +
  • Identity drifts because character anchors are vague or change between clauses.
  • The stream cuts unpredictably because the opening requests a montage.
  • Later directions erase the world because they restate a complete new scene instead of a delta.
02

Native realtime · prompt

Mid-Stream Scene Steering

Change one upcoming beat without accidentally rebooting the character, camera, or world.

Model
H3 Max Director
Mode
Prompt message
Use when
Sending a direction after a Director session is already configured and streaming.
RETIVAI RECIPE / 02
NEXT BEAT ONLY: Keep the same [CHARACTER], wardrobe, location, time, lighting, and camera language.

[CHARACTER] now [ONE CLEAR ACTION]. The camera [ONE CAMERA RESPONSE]. [ONE ENVIRONMENTAL REACTION].

Preserve every established object and spatial relationship. Continue naturally after this beat; do not end, reset, cut to a new place, add text, or introduce a new person.

Variables

[CHARACTER]
The existing character name or role already established in configure.
[ONE CLEAR ACTION]
One visible change achievable in the next segment.
[ONE CAMERA RESPONSE]
For example: slowly tracks left or holds a medium close-up.
[ONE ENVIRONMENTAL REACTION]
A small physical consequence that makes the action readable.

Continuity controls

  • Reference established facts instead of redescribing them with new adjectives.
  • Send one narrative beat per prompt version.
  • Leave the subject in motion so the next direction has a clean handoff.

Audio

If speech is needed, include one short exact line in quotes and keep the existing ambience unchanged.

Avoid

A new style bible, a location reset, several sequential actions, or camera directions that contradict the current shot.

Common failure modes +
  • The requested beat arrives late because the instruction contains a whole scene sequence.
  • Wardrobe or lighting shifts because the steering message unnecessarily redescribes them.
  • The result feels static because the instruction describes a pose instead of an action.
03

Native realtime · continuity

Persistent Character & World Lock

A compact identity sheet for sessions where the same person and place must survive many directions.

Model
H3 Max Director
Mode
Configure message
Use when
Launching a recurring host, character, or story world where continuity matters more than visual variety.
RETIVAI RECIPE / 03
CONTINUOUS WORLD BIBLE — [SERIES NAME]

1. [CHARACTER NAME]: [AGE RANGE], [FACE SHAPE / SKIN / DISTINCTIVE FEATURE].
2. Hair: [COLOR, LENGTH, SHAPE]. Never changes.
3. Wardrobe: [TOP], [BOTTOM], [FOOTWEAR], [ONE ACCESSORY]. Exact colors persist.
4. Scale and movement: [HEIGHT / BUILD], moves with [MOTION QUALITY].
5. Voice: [VOICE QUALITY]. Speaks only exact quoted lines.
6. Home location: [SINGLE LOCATION] with [THREE PERMANENT LANDMARKS].
7. Light: [ONE MOTIVATED LIGHT SOURCE] from [DIRECTION].
8. Camera: [ONE REPEATABLE CAMERA GRAMMAR].
9. World rule: [ONE PHYSICAL OR STORY RULE THAT NEVER CHANGES].
10. Current state: [WHAT THE CHARACTER WANTS AND WHAT OBJECTS ARE PRESENT].

OPENING: [CHARACTER NAME] is already [SIMPLE CONTINUOUS ACTION]. Preserve lines 1–10 across every upcoming segment. Later directions may change the action and story state, but never silently change identity, wardrobe, landmarks, camera grammar, or world rule.

Variables

[THREE PERMANENT LANDMARKS]
Visible spatial anchors that help the set remain recognizable.
[ONE REPEATABLE CAMERA GRAMMAR]
A single production language such as handheld documentary or locked studio dolly.
[CURRENT STATE]
Only facts that the next few beats must remember.

Continuity controls

  • Use numbered declarative facts rather than a long literary biography.
  • Update current state in steering prompts; keep immutable identity lines stable.
  • Prefer observable traits over personality labels the camera cannot show.

Audio

A stable voice description and exact quoted dialogue reduce the chance that speech style changes between beats.

Avoid

Brand-name characters, conflicting age/style descriptions, more than one home location, or a long list of invisible personality traits.

Common failure modes +
  • Character details compete because the sheet is too long.
  • The set rotates or rebuilds because landmarks are not spatially concrete.
  • Story progress stalls because immutable identity and mutable state are mixed together.
04

Interactive stream · moderation-ready

Audience-Requested Continuation

Translate a selected audience idea into one bounded, filmable continuation.

Model
H3 Max Director or H3 Max/Turbo I2V
Mode
Director prompt or next-clip prompt
Use when
A moderator or selection agent has chosen one audience message and needs to apply it without handing the model raw chat.
RETIVAI RECIPE / 04
AUDIENCE INTENT (already reviewed): [SANITIZED IDEA]

Continue from the exact current scene. Keep [CHARACTER ANCHORS], [LOCATION ANCHORS], current time, and established visual style unchanged.

In the next beat, [PROTAGONIST] attempts to [VISIBLE GOAL]. Show a clear beginning, physical consequence, and readable outcome: [SUCCESS / COMPLICATION / OPEN QUESTION]. The environment responds through [ONE CONCRETE EFFECT].

Camera: [SHOT AND MOVEMENT]. Audio: preserve the existing ambience; add [ONE SOUND CUE]. No new named people, brands, logos, interface text, or unrequested location change. End on active motion that supports another audience choice.

Variables

[SANITIZED IDEA]
The selected intent after policy filtering and prompt-injection removal.
[CHARACTER ANCHORS]
A short carry-forward of the established identity.
[VISIBLE GOAL]
An action the camera can judge, not an internal emotion.
[OUTCOME]
One controlled outcome category chosen by the game/show logic.

Continuity controls

  • Never paste raw audience text into the generation prompt.
  • Carry only the selected intent; system and world rules remain authoritative.
  • Store the accepted intent and resulting state change for the next beat.

Audio

Preserve the established sound bed and add one legible cue tied to the action. Avoid asking for several overlapping lines of dialogue.

Avoid

Raw usernames, prompt-injection text, copyrighted characters, public figures, brand marks, explicit harm, or an audience request that rewrites the system rules.

Common failure modes +
  • A raw comment hijacks the prompt because it bypassed sanitization.
  • The model invents a new setting because the continuation lacks current anchors.
  • The audience cannot tell whether its choice mattered because the outcome is internal or ambiguous.
05

Synthetic live · recurring format

Live AI Host & News Desk

Produce a short, clearly fictional desk segment with stable host identity and readable pacing.

Model
H3 Max or H3 Max Turbo
Mode
Text-to-video clip
Use when
Generating one segment for a buffered channel with a recurring original presenter.
RETIVAI RECIPE / 05
[00:00–00:02] Wide establishing shot of an original fictional studio called [PROGRAM NAME]. [HOST DESCRIPTION] sits behind [DESK ANCHOR]. The set uses [COLOR PALETTE] and one blank, abstract background panel with no text or logos. A restrained opening music sting resolves into quiet studio ambience.

[00:02–00:06] Slow camera push to a medium shot. The host looks into camera and says exactly: "[SHORT VERIFIED LINE]." Natural mouth movement, calm broadcast delivery, no other voices.

[00:06–00:10] The host turns toward [PHYSICAL PROP OR DEMO] and makes [ONE VISIBLE EXPLANATORY GESTURE]. The camera follows gently. End with the host still engaged, ready for the next chained shot.

Keep the host's face, hair, wardrobe, desk, lighting direction, and camera height identical to the supplied reference or previous final frame. This is a fictional program. Do not create a real network logo, real journalist, ticker, captions, lower third, watermark, or unsupported on-screen claim.

Variables

[HOST DESCRIPTION]
A compact, original character sheet reused in every segment.
[SHORT VERIFIED LINE]
One source-checked sentence short enough for the clip.
[PHYSICAL PROP OR DEMO]
A visible object that explains the topic without generated typography.

Continuity controls

  • Use the previous final frame as the next image-to-video input when available.
  • Keep the desk, camera height, key light, and wardrobe fixed across the program.
  • Put citations and precise text in the surrounding web UI, not inside generated pixels.

Audio

One quoted sentence plus quiet studio ambience. Use captions rendered by the web player if accessibility text is needed.

Avoid

Real network branding, real anchors, breaking-news urgency, fabricated statistics, dense on-screen text, or multiple speakers in one short clip.

Common failure modes +
  • Generated ticker text becomes unreadable or invents claims.
  • The host changes between clips because wardrobe and facial anchors are not repeated.
  • Long narration is truncated or garbled.
06

Buffered interactive · fixed world

Interactive Island Survival Show

Turn a voted survival choice into a clear beat while protecting a recurring cast and map.

Model
H3 Max or H3 Max Turbo
Mode
Image-to-video continuation
Use when
Generating the winning branch of a recurring survival show from the prior shot's final frame.
RETIVAI RECIPE / 06
SERIES LOCK: [SHOW NAME], an original grounded survival drama. Preserve the exact people already visible: [CAST ANCHORS]. Preserve [ISLAND ZONE], weather direction, time of day, wardrobe damage, carried tools, and established injuries.

WINNING AUDIENCE CHOICE: [SANITIZED CHOICE].

[00:00–00:03] Continue directly from the supplied final frame. [FOCAL CHARACTER] notices [IMMEDIATE PHYSICAL PROBLEM] and begins the chosen action. Handheld camera stays on the same side of the action.

[00:03–00:07] The action meets a believable obstacle: [OBSTACLE]. Show hands, tools, terrain, and cause-and-effect clearly; no unexplained object appears.

[00:07–00:10] Resolve this beat with [VISIBLE CONSEQUENCE]. End on a new, simple dilemma that can be expressed as two or three audience choices.

Audio: continuous wind, surf, and close practical sounds. Dialogue only if supplied here in quotes: "[OPTIONAL SHORT LINE]." No captions, logos, celebrity likenesses, fantasy rescue, sudden location change, or reset of damage and inventory.

Variables

[CAST ANCHORS]
Names plus one visible identity/wardrobe marker for each present character.
[ISLAND ZONE]
A stable map zone such as north beach camp or mangrove inlet.
[VISIBLE CONSEQUENCE]
A result that the next state reducer can record.

Continuity controls

  • Maintain a small state object for zone, cast present, health, inventory, weather, and unresolved threat.
  • Generate only the winning branch unless the experience explicitly budgets speculative branches.
  • Extract and retain the final frame plus a factual state update after every shot.

Audio

Repeat the same environmental sound bed across chained clips and limit dialogue to one short quoted line.

Avoid

Public-figure lookalikes, franchise characters, magically restored wardrobe, teleported props, or a consequence that exists only in narration.

Common failure modes +
  • A contestant disappears because too many characters are active in one short shot.
  • Inventory resets because carried tools are absent from the continuity state.
  • Voting feels cosmetic because the consequence does not change the next dilemma.
07

Interactive game · state-readable

First-Person AI Video Game Action

Stage one first-person action whose success or complication can be judged from the pixels.

Model
H3 Max or H3 Max Turbo
Mode
Image-to-video continuation
Use when
A game has selected a player action and needs the generated shot to expose a legible outcome.
RETIVAI RECIPE / 07
FIRST-PERSON CONTINUATION from the supplied final frame. Preserve [LOCATION], [LIGHTING], the player's visible [HANDS / GLOVES], held [ITEM], and all object positions.

PLAYER ACTION: [ONE PHYSICAL VERB + TARGET].
RULED OUTCOME: [SUCCESS / PARTIAL SUCCESS / FAILURE] because [PHYSICAL REASON].

[00:00–00:03] The player's hands begin the action in frame. Keep human-scale movement and a stable horizon.
[00:03–00:07] Show contact with the target and the exact physical reason for the ruled outcome.
[00:07–00:10] Hold long enough to read the consequence: [VISIBLE STATE CHANGE]. End with the target and relevant item still visible for the next decision.

Camera behaves like a body-mounted cinematic first-person view—natural head motion, no third-person cutaway, no floating HUD, no text, no extra limbs. Audio is location ambience plus synchronized contact sounds.

Variables

[RULED OUTCOME]
Chosen by game logic before generation, not improvised after the fact.
[PHYSICAL REASON]
A visible cause such as a rusted hinge or slipping rope.
[VISIBLE STATE CHANGE]
The frame evidence the game can use next.

Continuity controls

  • Keep the relevant item and target visible at the final frame.
  • Write the state change only after checking what actually appears on screen.
  • If a vision adjudicator disagrees with the planned result, preserve the visible outcome and flag the discrepancy.

Audio

Use synchronized Foley and existing ambience; avoid voiceover that announces an outcome the image does not show.

Avoid

Third-person cuts, game HUD text, invisible consequences, several chained actions, impossible camera movement, or extra hands.

Common failure modes +
  • The action is impossible to judge because contact happens off-screen.
  • A planned success appears as failure, creating a state mismatch.
  • The final frame hides the item needed for the next continuation.
08

Education · single concept

Five-Second Classroom Explainer

Illustrate one concrete learning idea in a short beat that can join a longer lesson runway.

Model
H3 Max or H3 Max Turbo
Mode
Text-to-video clip
Use when
A lesson planner has already reduced an explanation to one visualizable subject.
RETIVAI RECIPE / 08
STYLE LOCK: [ORIGINAL ILLUSTRATION STYLE], using [LIMITED COLOR PALETTE], stable line weight, clean shapes, and no written words.

SCIENTIFIC SUBJECT: [ONE CONCRETE CONCEPT].
TEACHING INTENT: Make the viewer understand [ONE OBSERVABLE RELATIONSHIP].

Create exactly one five-second 16:9 illustrated demonstration. Begin with [INITIAL STATE]. Animate one change: [PROCESS]. End on [FINAL STATE] and hold it briefly so the relationship is readable.

Composition stays centered with one dominant subject and one supporting element. Use scale, motion, color contrast, and spatial relation instead of labels. No presenter, equations, captions, logo, decorative text, unrelated objects, or second concept.

Audio: one soft transition sound synchronized to the change; no speech.

Variables

[ONE CONCRETE CONCEPT]
For example: warm air rises, not the whole topic of weather.
[ONE OBSERVABLE RELATIONSHIP]
The before/after relationship the image must teach.
[PROCESS]
One motion that expresses the concept.

Continuity controls

  • Reuse the same style lock and palette across every lesson beat.
  • Let the lesson player add accurate captions outside the generated video.
  • Plan all lesson beats together, then render each visual separately.

Audio

Keep generated audio minimal; narration and accessible captions are more reliable when the player owns them.

Avoid

Dense diagrams, written labels, more than one scientific process, anthropomorphic claims, or visually unverifiable facts.

Common failure modes +
  • The concept becomes decorative because the physical relation is not visible.
  • Generated labels are misspelled.
  • Style changes from one lesson beat to the next.
09

Reliability · buffer recovery

Safe Filler & Recovery Scene

Generate a low-risk holding beat that buys time without silently advancing the story.

Model
H3 Max or H3 Max Turbo
Mode
Image-to-video continuation
Use when
The next planned segment failed, moderation is unresolved, or the generation runway is below its safety threshold.
RETIVAI RECIPE / 09
RECOVERY BEAT from the supplied final frame. Preserve every established character, object, location, wardrobe detail, light direction, and camera style.

Do not advance the plot, introduce information, solve a problem, move to a new place, add a person, or change inventory.

For [DURATION] seconds, show [LOW-RISK CONTINUOUS ACTION]: subtle breathing, fabric and hair responding to the environment, [AMBIENT MOTION], and a gentle [CAMERA MOTION]. Keep the focal character oriented toward the current unresolved situation.

Audio continues seamlessly with [EXISTING AMBIENCE] and one quiet environmental detail. No dialogue, captions, signs, logos, sudden event, threat escalation, or fade to black. End in a composition that can accept the original next action.

Variables

[LOW-RISK CONTINUOUS ACTION]
A neutral behavior that does not create new state.
[AMBIENT MOTION]
Rain, leaves, dust, steam, or another established background motion.
[DURATION]
A short segment matched to the pipeline's minimum useful buffer.

Continuity controls

  • Tag filler separately so it never enters the factual story summary.
  • Pre-render a rights-safe generic fallback for total provider failure.
  • Resume only from the actual final filler frame if it becomes part of the visible stream.

Audio

Carry the current ambience without a musical climax or line of dialogue that implies a new event.

Avoid

New props, reveals, plot progress, camera cuts, black frames, logos, or a fallback that falsely looks like live output.

Common failure modes +
  • The filler accidentally changes story state.
  • Repeated filler becomes visibly identical and exposes the loop.
  • A hard cut back to the main scene breaks continuity.

Editorial method

How these recipes
were made.

01

Read primary structures

We reviewed fal's Director message schema and MiniMax's official H3 prompt-writing materials, then inspected public products and repositories that expose their realtime loops.

02

Write from scratch

Prompt text is original RetivAI editorial work. Public collections inform categories and validation ideas, not copy-and-paste inventory.

03

Label the evidence

“Official structure” means the message shape is documented. “Source-tested pattern” means a public project's source describes the pattern. Neither means RetivAI generated the output.

04

Revise after real testing

When RetivAI funds generation tests, the result, model version, settings, date, failed attempts, and media evidence will be recorded before any testing label changes.

Need the product loop too? A good prompt cannot repair missing state, buffering, moderation, or spend limits.

Open product templates