🎬 LTX 2.5 Video Prompt Agent Skill

A Hermes Agent skill that transforms rough video ideas + reference images/audio into production-grade LTX 2.5 (Lightricks) video generation prompts.

Unlike structured-format skills (e.g., MiniMax H3), LTX 2.5 prompting is pure natural language — one flowing cinematic paragraph that reads like a shot description for a cinematographer. No JSON, no field labels, no tags. Just prose.

What It Does

When you describe a video concept (with or without attached media), this skill:

1. Detects the mode — single-shot, multishot (NEW in 2.5), or screenplay-style
2. Creatively enhances your idea across 6 cinematic dimensions
3. Outputs a natural-language prompt — ready to paste into ComfyUI or LTX Studio

LTX 2.5 Modes

⭐ Native multishot is the headline feature of LTX 2.5. Previous versions only produced a single continuous shot. Now you can generate 2–4 connected shots — holding character identity, environment, lighting, and voice across cuts — all from one prompt.

The 6-Part Prompt Structure

Every prompt integrates these elements in order (models weight earliest tokens more heavily):

1. 🎥 Establish the Shot — Cinematography terms, shot scale, angle, style markers
2. 💡 Set the Scene — Lighting, color palette, textures, atmosphere
3. 🏃 Describe the Action — Core action as a flowing sequence, present tense
4. 👤 Define Character(s) — Physical specifics, wardrobe, emotion via physical cues
5. 📷 Camera Movement — How and when the camera moves, relative to subject
6. 🔊 Describe the Audio — Ambient sound, action SFX, music, dialogue

Important: This structure is a mental model for the writer. The output is pure prose — the model never sees field labels or numbered sections.

Installation

For Hermes Agent Users

Copy the skill folder to your Hermes skills directory:

# Linux/macOS
cp -r ltx-2-5-prompter ~/.hermes/skills/creative/

# Windows
xcopy ltx-2-5-prompter %LOCALAPPDATA%\hermes\skills\creative\ /E /I

Restart Hermes Agent. The skill auto-triggers when you mention LTX, attach media, and describe a video idea.

For Other Agent Frameworks (Claude, Codex, etc.)

The SKILL.md and reference files are standard Markdown — load them as system context or project rules in any agent that supports file-based instructions.

File Structure

ltx-2-5-prompter/
├── README.md                            # You are here
├── SKILL.md                             # Main skill — workflow, rules, verification
└── references/
    ├── single-shot-format.md            # Default mode — one continuous take
    ├── multishot-format.md              # NEW 2.5 feature — connected cuts in one prompt
    ├── screenplay-format.md             # Dialogue-heavy scenes with character cues
    ├── creative-examples.md             # Weak→strong transformations + advanced patterns
    └── ltx-vocabulary.md                # 100+ categorized cinematic terms

Usage Examples

Single-Shot (Text-to-Video)

💬 "A golden retriever running through a sunny meadow"

→ The skill enhances this into:

Wide tracking shot of a golden retriever sprinting through a sunlit wildflower meadow, golden hour light catching the fur, lens flares from the low sun, camera follows at ground level alongside the dog, shallow depth of field with wildflowers blurring in the foreground. The sound of padded footsteps on grass, distant birdsong, and a warm breeze rustling the flowers.

Multi-Shot (NEW in 2.5)

💬 "A woman in a raincoat waiting for someone at a rainy intersection, 
    then we see the person arriving — multishot"

→ The skill produces a chronological paragraph with named cuts, audio continuity, and character re-identification at each transition. 2–4 shots, each with a clear job (establish → detail → reaction).

Screenplay-Style (Dialogue Scene)

💬 "A news reporter discovering oil has been found in a small town, 
    live broadcast format"

→ The skill formats the scene with INT./EXT. headers, character cues, physical acting directions between dialogue beats, and camera/audio notes woven throughout.

Weak → Strong Transformation

What’s New in LTX 2.5 (vs 2.3)

Key Prompting Principles

1. Natural language, not tags — LTX reads prose like a cinematographer reads a shot description
2. One action per shot — never cram multiple actions; split across shots or generations
3. Physical cues, not emotional labels — “eyes widen, voice cracking” not “sad”
4. Named references, not vague qualifiers — “shot on 35mm, documentary look” not “cinematic”
5. Match length to duration — short prompts for long videos leave the model without direction
6. One coherent light source — mixed lighting confuses scene interpretation
7. Always describe audio — LTX 2.5 generates synced audio natively
8. Present tense throughout — “walks” not “walked”

Multishot Transition Vocabulary

At every cut, state: new framing, who’s in frame (re-identified), and audio continuity (“the score continues across the cut” or “the music drops; only wind remains”).

Technical Constraints (LTX 2.5)

• Frame count: num_frames % 8 == 1 (1, 9, 17, 25, …, 121, …)
• Resolution: Width and height divisible by 32
• Distilled model: Fixed 8-step schedule, CFG=1
• Auto-duration: Omit --num-frames to let the duration head predict clip length
• VRAM: 16GB minimum (int8 convrot), 24GB+ recommended (bf16)
• Spatial upscaler: LTX 2.3 spatial upscaler still required for Stage 2

Sources & References

• LTX 2.5 Model Page — Official features and benchmarks
• LTX 2.5 HuggingFace — Model card, weights, code samples
• LTX 2.5 Prompt Guide — Official prompting guide
• LTX 2.3 Prompt Guide — Previous version reference
• AI Video Prompt Guide — General 6-part structure

License

MIT — use freely. The LTX 2.5 model weights themselves are under the LTX-2.x Community License (free for organizations under $10M ARR).

Author

Benji AI Playground — YouTube · GitHub

In-depth technical AI video content. H3, LTX, SeeDance pipelines and workflows.