Skip to content
Back to search

h3-video-prompt-enhancer

"Use when making MiniMax H3 video prompts from media + ideas."

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/benjiyaya--minimax-h3-prompt-agentskill/badge.svg)](/listings/benjiyaya--minimax-h3-prompt-agentskill)

HTML

<a href="/listings/benjiyaya--minimax-h3-prompt-agentskill"><img src="/api/listings/benjiyaya--minimax-h3-prompt-agentskill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

For Hermes Agent Users

Copy the skill folder to your Hermes skills directory:

# Linux/macOS
cp -r h3-video-prompt-enhancer ~/.hermes/skills/creative/

# Windows
xcopy h3-video-prompt-enhancer %LOCALAPPDATA%\hermes\skills\creative\ /E /I

Restart Hermes Agent. The skill auto-triggers when you 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.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 216 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: h3-video-prompt-enhancer
  3. description: "Use when making MiniMax H3 video prompts from media + ideas."
  4. version: 1.0.0
  5. author: Hermes Agent
  6. license: MIT
  7. metadata:
  8. hermes:
  9. tags: [video, prompt-engineering, minimax-h3, comfyui, text-to-video, image-to-video, creative]
  10. related_skills: [comfyui]
  11. ---
  12. # H3 Video Prompt Enhancer
  13. ## Overview
  14. Transform a user's rough video idea + attached assets into a production-grade MiniMax H3 video generation prompt. The skill handles two H3 generation modes:
  15. - **Ref2VA** — when the user provides reference assets (character sheets, style images, reference videos, voice samples). Outputs a 6-section structured prompt. Load `references/ref2va-format.md` for the full spec.
  16. - **Base MultiShot** (T2VA / I2VA / FL2VA / L2VA) — for text-to-video or image-anchored generation. Outputs a timed multi-shot prompt. Load `references/base-multishot-format.md` for the full spec.
  17. The skill's distinctive value: it doesn't just format-comply — it **creatively enhances** the brief with professional-grade cinematic detail (camera aesthetics, visual texture, lighting design, pacing arcs, spatial choreography, continuity tracking) before mapping everything into the exact H3 output format. Load `references/creative-showcase.md` for quality benchmarks and pattern examples from advanced long-form prompts.
  18. ## When to Use
  19. **Trigger when the user:**
  20. - Attaches 1+ images/videos/audio files and describes a video they want to create
  21. - Says "make a video prompt" / "enhance this for H3" / "write a Ref2VA prompt" / "img2vid" / "txt2vid"
  22. - Provides a video concept and wants it structured for MiniMax H3 generation
  23. - Mentions H3, MiniMax, Ref2VA, ComfyUI video generation
  24. - Pastes a creative brief (CAMERA/LOOK/STYLE format) and wants it converted to H3 format
  25. **Don't use for:**
  26. - Non-H3 video models (Sora, Runway, Kling, etc.) — the output format is H3-specific
  27. - Pure image generation prompts
  28. - Video editing tasks that don't involve new generation
  29. ## Step 0: Classify the Mode
  30. Determine the H3 mode from what the user attached and stated:
  31. | User provides | Mode | Format reference |
  32. |---|---|---|
  33. | Reference images/videos/audio (character sheets, style refs, voice clips) | **Ref2VA** | `references/ref2va-format.md` |
  34. | Nothing — just a text idea | **T2VA** (text-to-video) | `references/base-multishot-format.md` |
  35. | 1 image as the first frame | **I2VA** (image-to-video) | `references/base-multishot-format.md` |
  36. | 2 images (first frame + last frame) | **FL2VA** (first-last-to-video) | `references/base-multishot-format.md` |
  37. | 1 image as the last frame only | **L2VA** (last-frame-to-video) | `references/base-multishot-format.md` |
  38. **Key distinction:** An image used as a *first/last frame anchor* = I2VA/FL2VA/L2VA. An image used as a *character/style reference* (not a frame position) = Ref2VA. When ambiguous, ask the user: "Is this image a frame anchor (first/last frame of the video) or a reference (character/style/scene)?"
  39. ## Step 1: Gather Parameters
  40. Confirm these before enhancing (ask if missing, but proceed if the idea is clear enough):
  41. - **duration_s**: 4–15 seconds (integer). Default to 8 if unspecified.
  42. - **aspect ratio**: 16:9, 9:16, 1:1, 4:3, 21:9. Default to 16:9.
  43. - **shot count**: Let the planner decide, or respect user's explicit count. Budget: 4–6s → 1–2 shots; 7–10s → 2–3 shots; 11–15s → 3–5 shots.
  44. - **asset inventory**: What each attached file is and its role. Numbering per modality defines labels: Image k → `<Picture k>` or `<Subject N>`; Video k → `<Video k>`; Audio k → `<Audio k>`.
  45. ## Step 2: Creative Enhancement
  46. This is the core value-add. Take the user's idea and enrich it across seven dimensions. The goal: produce a prompt with the depth and cinematic intelligence of a professional storyboard. Consult `references/creative-showcase.md` for full examples of each pattern.
  47. ### Enhancement Dimensions
  48. **1. Camera Identity** — Assign a distinct camera aesthetic that matches the scene's tone:
  49. - Physical type: handheld, tripod, drone, steadicam, dolly, security cam, dashcam, POV, arc shot
  50. - Imperfections to preserve (when stylistically appropriate): hand tremor, autofocus hunting, exposure fluctuation, lens flare, motion blur, awkward zoom
  51. - Format/aesthetic hint: 16mm film, DV tape, digital clean, anamorphic, vintage camcorder, broadcast
  52. **2. Visual Texture (LOOK)** — Define the image quality and color science:
  53. - Grain/noise: film grain, electronic noise, clean digital, VHS tracking, subtle blur
  54. - Color palette: warm/cool, saturated/desaturated, high/low contrast, natural skin tones
  55. - Lighting design: natural, studio, neon, golden hour, mixed sources, practical lights
  56. - Lighting transitions: if locations change across shots, describe how the light shifts
  57. **3. Pacing Arc** — Plan the energy progression across the full duration:
  58. - Build patterns: quiet→energetic, tense→release, slow build→explosive peak→settle, steady rhythm
  59. - Cut rhythm: accelerating cuts toward climax, slow contemplative holds, musical cutting on beats
  60. - Match the pacing arc to the emotional intent of the brief
  61. **4. Character Detail** — Flesh out every on-screen person with specificity:
  62. - Physical: age range, build, hair color/style, skin tone, distinctive features (scars, freckles, heterochromia)
  63. - Wardrobe: specific garments with colors, materials, textures, accessories — note changes across locations or time
  64. - Visual signature: a recurring color or visual element that makes the character instantly recognizable in every shot (e.g., "electric purple energy trails", "glossy teal jacket reflections", "always wears a red scarf")
  65. - Coverage note: describe outfits fully — avoid implying revealing clothing unless the user explicitly requests it
  66. **5. Spatial Geography** — For action sequences or multi-location videos:
  67. - Screen direction: who enters from where, movement vectors (Left→Right, Deep→Front, foreground↔background)
  68. - Key action moments: the 2–3 critical motion beats that define the sequence
  69. - Environmental layout: what's in the space, how it's lit, reflective surfaces, depth
  70. **6. Continuity Progression** — Track what changes across shots so the video feels coherent:
  71. - Physical state: damage accumulates, hair gets messier, clothes get wet/torn/dusty
  72. - Environmental: props move, lights flicker, weather shifts, debris scatters
  73. - Emotional: expressions and body language evolve naturally across the timeline
  74. **7. Sound Design Plan** — Map the full audio landscape:
  75. - Ambience: room tone, environmental atmosphere, background chatter, traffic, wind
  76. - Physical action sounds: footsteps, impacts, fabric rustling, door creaks, liquid pouring
  77. - Non-diegetic score: instrumentation, tempo, rhythm, dynamic changes (goes in `non_diegetic_music` field)
  78. - Diegetic music: source-visible music — radio, speaker, live performance (goes in shot description)
  79. - Dialogue vs. voiceover: clearly mark which is spoken on-camera vs. off-screen narration
  80. ### Per-Shot Quality Bar
  81. Every shot in the storyboard must specify:
  82. - **Composition**: framing (wide, medium, close-up, extreme close-up, macro), angle (eye-level, low, high, overhead, Dutch)
  83. - **Camera motion**: type + amplitude + speed in natural English (e.g., "The camera pushes in with small amplitude at slow speed")
  84. - **Subject action**: exactly ONE dominant action per shot — never cram multiple actions
  85. - **Environment/lighting**: what's visible, how it's lit, time of day cues
  86. - **Sound cue**: what's audible in this specific moment
  87. - **Reference labels** (Ref2VA only): where referenced content appears or takes effect
  88. ## Step 3: Format and Output
  89. Load the appropriate format reference and produce the final H3 prompt.
  90. **For Ref2VA** → Load `references/ref2va-format.md`. Output exactly 6 sections in order:
  91. 1. `subject_definitions:` — one line per tracked item
  92. 2. `summary:` — task-type prefix + one paragraph
  93. 3. `retention_analysis:` — fidelity markers per label
  94. 4. `detailed_description:` — 350–500 words, opens with style, then `[Shot N]` timeline
  95. 5. `overall_soundscape:` — 1–4 sentences
  96. 6. `non_diegetic_music:` — 1–3 sentences or N/A
  97. **For Base MultiShot** → Load `references/base-multishot-format.md`. Output:
  98. 1. Instruction line (I2VA/FL2VA/L2VA only — exact template from reference)
  99. 2. `integrated_multimodal_description:` — timed multi-shot timeline
  100. 3. `overall_soundscape:` — 1–4 sentences
  101. 4. `non_diegetic_music:` — 1–3 sentences or N/A
  102. ### Critical Format Rules (both modes)
  103. - Output ONLY the specified fields — no preamble, no explanation, no markdown fences, no commentary
  104. - Write everything in English. Exceptions: dialogue/lyrics inside `<d>` tags and visible on-screen text keep their original language verbatim
  105. - Timestamps: `[Shot 1]` has NO timestamp and opens with style + initial composition. Later shots: `[Shot N] At MM:SS.mmm, the camera cuts to ...` with strictly increasing times within duration
  106. - A cut must add NEW information (subject, space, state, viewpoint, time). If only distance/angle changes, use camera motion instead of a cut
  107. - Camera motion = motion type + amplitude + speed as natural English action, not stacked labels
  108. - Dialogue format: identifying phrase + speaker ID + delivery OUTSIDE `<d>`; inside `<d>` ONLY language tag + exact words: `<d>[English] Wait for us!</d>`
  109. - Preserve the user's dialogue verbatim — never translate, rewrite, or paraphrase
  110. - Voiceover: exact phrase "says in an off-screen voiceover" + immediately state "while his/her lips remain completely closed"
  111. - On-screen text (signs, neon, subtitles): English double quotation marks, verbatim, no translation
  112. - Avoid named third-party IP, real celebrities, trademarked characters — describe them generically
  113. ### Camera Motion Vocabulary
  114. Types: Zoom In/Out, Push In/Pull Out, Pan L/R, Truck L/R, Tilt Up/Down, Pedestal Up/Down, Arc Shot, Tracking Shot, Static Shot, Shake Slightly/Strongly, POV, Roll CW/CCW.
  115. Amplitude: "with small amplitude" / "with large amplitude" (omit when medium).
  116. Speed: "at slow speed" / "at fast speed" (omit when normal).
  117. ## Step 4: Present to User
  118. After generating the H3 prompt:
  119. 1. Present the full prompt in a code block so it can be copied directly
  120. 2. State which mode was detected and why
  121. 3. Flag any assumptions made (e.g., "Assumed 8s duration and 16:9 — adjust if needed")
  122. 4. Offer to refine specific aspects: camera aesthetic, pacing, character detail, shot count, sound design
  123. ## Creative Patterns from Showcase
  124. The user maintains advanced long-form prompt examples representing the quality bar. Load `references/creative-showcase.md` for full examples. Key transferable patterns:
  125. **Long-form Storytelling** (montage, day-in-the-life, narrative):
  126. - Explicit CAMERA/LOOK/STYLE blocks defining the aesthetic before the storyboard
  127. - Outfit and location progression across time, with lighting transitions
  128. - Voiceover carrying emotional narrative over non-synchronized visuals
  129. - Pace building from quiet opening to high-energy finale
  130. - Per-shot storyboard with timing, action, and VO lines
  131. **Action Choreography** (combat, chase, sports):
  132. - Per-character color lock — each character has a signature color for their effects/trails/reflections
  133. - Explicit spatial layout with screen directions (Deep→Front, Left→Right)
  134. - Action vectors — the key motion moment that defines each shot
  135. - Progressive continuity — damage accumulates, hair gets windblown, dust/cracks appear
  136. - Strict shot count enforcement — state the count and hold it
  137. - Environmental reactivity — holograms flicker, floors reflect, walls crack in response to action
  138. ## Common Pitfalls
  139. 1. **Treating Ref2VA reference images as frame anchors.** A character sheet or style reference is a `<Subject>`, not a `<Picture>`. Only use `<Picture N>` standalone when the image IS a concrete frame position (first frame, keyframe, last frame). When in doubt, cite the image inside the relevant `<Subject N>` line.
  140. 2. **Cramming multiple actions into one shot.** One dominant action per shot — this is a hard H3 constraint. If the brief describes sequential actions, split them across multiple shots with cuts.
  141. 3. **Forgetting camera motion.** Every shot needs an explicit camera movement specification — even "Static Shot" if the camera doesn't move. Omitting it leaves the model guessing.
  142. 4. **Inconsistent character identity across shots.** Repeat identity anchors (hair, clothing, key props) in every shot, phrased freshly but consistently. If hair gets messy in shot 3, it stays messy in shot 4.
  143. 5. **Wrong shot count for duration.** Respect the budget: 4–6s → 1–2 shots; 7–10s → 2–3 shots; 11–15s → 3–5 shots. Don't plan 5 shots for a 5-second video.
  144. 6. **Mixing diegetic and non-diegetic music.** Music audible to characters (radio, live performance, phone speaker) goes in the shot description. Background score the characters can't hear goes in `non_diegetic_music`. Never put the same music in both.
  145. 7. **Inventing reference labels (Ref2VA).** Never create labels beyond those defined in `subject_definitions`. `<Subject 3>` means the same thing in every section where it appears. Different indices for different modalities are numbered independently.
  146. 8. **Translating or rewriting dialogue.** Preserve the user's exact words inside `<d>` tags — including punctuation, hesitations, and language. Never translate to English if the user wrote in another language.
  147. 9. **Skipping the style opener.** The `detailed_description` (Ref2VA) or `integrated_multimodal_description` (Base) MUST open with 1–2 sentences of overall style BEFORE `[Shot 1]`. Don't jump straight into the first shot.
  148. 10. **Flat timestamps.** `[Shot 1]` never has a timestamp. Every subsequent shot needs `At MM:SS.mmm` with strictly increasing times. Convert duration to seconds with proper formatting (8s → 8.00 for the instruction line).
  149. ## Verification Checklist
  150. - [ ] Mode correctly detected from attached assets and user intent
  151. - [ ] All seven enhancement dimensions applied (camera, look, pacing, character, spatial, continuity, sound)
  152. - [ ] Output contains ONLY the required H3 fields — no preamble, no markdown fences, no extra commentary
  153. - [ ] Timestamps are strictly increasing and fall within duration_s
  154. - [ ] Exactly ONE dominant action per shot
  155. - [ ] Camera motion specified for every shot (type + amplitude + speed)
  156. - [ ] Character identity consistent across all shots (repeat anchors every shot)
  157. - [ ] Dialogue preserved verbatim in `<d>[Language] ...</d>` format
  158. - [ ] Style opener present before `[Shot 1]`
  159. - [ ] Reference labels (Ref2VA only) used consistently and never invented beyond subject_definitions
  160. - [ ] Duration (4–15s) and aspect ratio specified or assumed-and-flagged
  161. - [ ] No named IP, celebrities, or trademarked character names
  162. - [ ] Diegetic music in shot descriptions, non-diegetic score in `non_diegetic_music`
README.md · README · 122 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # 🎬 MiniMax H3 Video Prompt Agent Skill
  2. A [Hermes Agent](https://github.com/NousResearch/hermes-agent) skill that transforms rough video ideas + reference images/audio into production-grade **MiniMax H3** video generation prompts.
  3. ## What It Does
  4. When you attach images, video, or audio and describe a video concept, this skill:
  5. 1. **Detects the H3 mode** automatically (Ref2VA, T2VA, I2VA, FL2VA, L2VA)
  6. 2. **Creatively enhances** your idea across 7 cinematic dimensions
  7. 3. **Outputs the exact H3 format** — ready to paste into ComfyUI
  8. ## H3 Modes Supported
  9. | Mode | Use Case | Output Format |
  10. |---|---|---|
  11. | **Ref2VA** | Character/style/voice reference assets | 6-section structured prompt |
  12. | **T2VA** | Text-to-video (no images) | Timed multi-shot timeline |
  13. | **I2VA** | 1 image as first frame | Instruction line + timeline |
  14. | **FL2VA** | 2 images (first + last frame) | Instruction line + timeline |
  15. | **L2VA** | 1 image as last frame | Instruction line + timeline |
  16. ## Creative Enhancement Dimensions
  17. Every prompt is enriched across:
  18. - 🎥 **Camera Identity** — Physical camera type, imperfections, format aesthetic
  19. - 🎨 **Visual Texture** — Grain, color palette, lighting design, exposure behavior
  20. - ⚡ **Pacing Arc** — Energy progression, cut rhythm, build patterns
  21. - 👤 **Character Detail** — Physical features, wardrobe, visual signature colors
  22. - 🧭 **Spatial Geography** — Screen directions, action vectors, environmental layout
  23. - 🔄 **Continuity Progression** — State tracking, damage accumulation, emotional arcs
  24. - 🔊 **Sound Design** — Ambience, SFX, diegetic/non-diegetic music mapping
  25. ## Installation
  26. ### For Hermes Agent Users
  27. Copy the skill folder to your Hermes skills directory:
  28. ```bash
  29. # Linux/macOS
  30. cp -r h3-video-prompt-enhancer ~/.hermes/skills/creative/
  31. # Windows
  32. xcopy h3-video-prompt-enhancer %LOCALAPPDATA%\hermes\skills\creative\ /E /I
  33. ```
  34. Restart Hermes Agent. The skill auto-triggers when you attach media and describe a video idea.
  35. ### For Other Agent Frameworks (Claude, Codex, etc.)
  36. 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.
  37. ## File Structure
  38. ```
  39. h3-video-prompt-enhancer/
  40. ├── README.md # You are here
  41. ├── SKILL.md # Main skill — workflow, rules, verification
  42. └── references/
  43. ├── ref2va-format.md # Full Ref2VA 6-section format spec
  44. ├── base-multishot-format.md # T2VA / I2VA / FL2VA / L2VA format spec
  45. └── creative-showcase.md # Advanced long-form prompt patterns & benchmarks
  46. ```
  47. ## Usage Examples
  48. ### Ref2VA (Character References)
  49. ```
  50. 📎 Attach: character_sheet.png, voice_sample.mp3
  51. 💬 "A woman matching this character walks through a neon-lit Tokyo street at night,
  52. stops at a ramen stand, says hi to the cook"
  53. ```
  54. → Outputs a 6-section structured prompt with `subject_definitions`, `summary`, `retention_analysis`, `detailed_description`, `overall_soundscape`, `non_diegetic_music`.
  55. ### I2VA (Image-to-Video)
  56. ```
  57. 📎 Attach: first_frame.png
  58. 💬 "A baker opening his shop before sunrise, proud of the first loaf"
  59. ```
  60. → Outputs a timed multi-shot timeline anchored to the first-frame image.
  61. ### FL2VA (First + Last Frame)
  62. ```
  63. 📎 Attach: first_frame.png, last_frame.png
  64. 💬 "Day to night transition over a city skyline"
  65. ```
  66. → Outputs a continuous interpolation prompt from first to last frame.
  67. ## Key Format Rules
  68. - Output is **raw structured text** — no markdown fences, no preamble
  69. - All timestamps use `MM:SS.mmm` format, strictly increasing
  70. - One dominant action per shot — never cram multiple actions
  71. - Camera motion written as natural English (e.g., *"The camera pushes in with small amplitude at slow speed"*)
  72. - Dialogue preserved verbatim inside `<d>[Language] ...</d>` tags
  73. - Reference labels (`<Subject N>`, `<Picture N>`, `<Video N>`, `<Audio N>`) stay consistent across all sections
  74. ## ComfyUI Integration
  75. ```
  76. LLM Node (with system prompt) → text output → H3 Ref2VA / H3 Base node prompt input
  77. ```
  78. Keep the ASSETS list order identical to the physical wiring order of the reference inputs — labels map positionally (Image 1 → first image socket, etc.).
  79. ## Provenance
  80. Built for the [MiniMax H3](https://github.com/MiniMax-AI) omni-modal video model. The Ref2VA and Base MultiShot format specs are derived from MiniMax's official H3-Context-IR preprocessor documentation.
  81. ## License
  82. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.