| name | helios-prompts |
| description | Write effective prompt sequences for Reactor's Helios model, a real-time video frame generation API. Use when creating Helios prompts, video generation prompts, or when the user mentions Helios or Reactor. Covers start prompts, follow-ups, sequence planning, camera anchoring, lighting consistency, and progressive state tracking. |
Helios Prompt Writing
Helios generates video sequences through multi-prompt continuation. Each prompt builds on the previous frame state with controlled variations while maintaining visual coherence.
Start Prompt Structure
Every start prompt follows this exact order:
[SUBJECT ESTABLISHMENT] 2-3 sentences. Specific physical details, clothing, posture, name.
[ENVIRONMENT DESCRIPTION] 2-3 sentences. Setting, props, background elements. Concrete details.
[LIGHTING CONDITIONS] 1 sentence. Source + quality. This becomes the "lighting anchor."
[ATMOSPHERE/MOOD] 1 sentence. Emotional tone.
[STYLE (optional)] 1 sentence. Only if non-default (VHS, anime, vintage).
[CAMERA INSTRUCTION] 1 sentence. Shot type + angle + focus. IDENTICAL in every prompt.
Subject: Not "a lion" but "a majestic lion named Leo." Include name, posture, clothing, key features.
Environment: Not "a table" but "a wooden dining table." Include specific props and background.
Lighting anchor: Be specific โ "warm afternoon sunlight streaming through linen curtains," not just "daylight." This exact description repeats in every follow-up.
Camera instruction: This line is copied VERBATIM to every prompt in the sequence. Lock shot type, angle, and focus.
Follow-Up Prompt Structure
Each follow-up = Base State + ONE New Action.
[SUBJECT REESTABLISHMENT] 1-2 sentences. Abbreviated restate of position/details.
[NEW ACTION] 1-2 sentences. Temporal connector + single action.
[ENVIRONMENTAL INTERACTION] 1-2 sentences. How environment reacts (steam, smoke, reflections).
[LIGHTING ANCHOR] 1 sentence. Copied from start, minor wording variations OK.
[ATMOSPHERE (if changed)] 1 sentence. Only if mood shifted.
[CAMERA INSTRUCTION] IDENTICAL to start prompt. Not one word different.
Reestablishment pattern: "The [subject] remains/sits/stands [location], [key visuals]. [Clothing reminder]."
Temporal connectors: "Suddenly...", "Now...", "As [subject]...", "[Subject] suddenly..."
Action categories (pick ONE per prompt):
- Micro-gestures: reaching, turning head, leaning, raising hand
- Object interaction: picking up, setting down, holding
- Facial expression: smiling, eyes widening, mouth opening
- Environmental trigger: breeze starting, something appearing
Environmental reactions โ always describe how the world responds:
- Objects affected by movement
- Light interacting with surfaces
- Particles (steam, smoke, petals, sparks)
- Reflections changing
Progressive State Tracking
Track cumulative changes across prompts. Objects and subjects evolve:
- Candles: lit โ flickering โ extinguished โ smoking โ smoke dissipated
- Flower: closed bud โ separating โ half-open โ fully bloomed
- Party hat: slightly tilted โ crooked โ slipped to one side
- Petals: scattered โ swirling โ covering blanket โ landing on face
Each follow-up must reflect the accumulated state from all previous prompts.
Sequence Planning
Optimal length: 4-5 prompts.
| Prompt | Role |
|---|
| 0 (Start) | Establish scene |
| 1-2 | Build action |
| 3 | Climax / payoff |
| 4 (optional) | Resolution / settling |
Avoid: >6 prompts (coherence degrades), single follow-ups (no arc), unnecessary intermediate steps.
Common Patterns
Choose one to structure the sequence. See references/patterns.md for detailed examples.
- Environmental Interaction โ Subject experiences active environment (rain, wind, petals, fireworks). Calm โ activate โ interact โ peak โ settle.
- Object Reveal/Transformation โ Something hidden is revealed or changes state. Initial โ transition โ reveal โ reaction โ enjoyment.
- Escalating Intensity โ Single element increases in quantity/power. Calm โ first โ multiple โ peak โ sustained.
- Character Action Sequence โ Subject performs related micro-actions. Rest โ prepare โ act โ result โ settle.
Critical Rules
DO:
- Keep lighting consistent (only minor wording variations)
- Reestablish subject (abbreviated) in every follow-up
- Introduce exactly ONE new action per follow-up
- Describe environmental reactions (steam, smoke, reflections, particles)
- Use specific object descriptions ("a plastic cake server with a bright green handle")
- Copy camera instruction verbatim to every prompt
- Use temporal connectors to signal changes
- Track progressive state changes across prompts
- Introduce objects with action verbs (picks up, pulls out, grabs)
DON'T:
- Change lighting dramatically (no dawn โ noon transitions)
- Stack multiple new actions in one prompt
- Modify the camera instruction between prompts
- Let subjects interact with dangerous elements in motion
- Introduce objects without describing how they appear
- Skip environmental reaction descriptions
Checklist
Before finalizing any sequence, verify:
