AI Media Generation
AI video and image prompt engineering — video generation, image generation, product video, social media video, cinematic direction, animation, Seedance, Higgsfield, GPT-Image-2. Applies 15 Maxim-autho
SKILL.md.claude/skills/ai-media-generation/SKILL.md
GitHubAI Media Generation — Domain Dispatcher
Maxim's skill domain for AI video and image prompt engineering. Combines 15 Maxim-authored cinematic styles (behavioral-moat-framed, see cinematic-styles/) with external technical references (20 Higgsfield sub-skills) — total 35 skills, 15 of which are Maxim's original IP with defensible behavioral-framework engineering.
Activation
This domain activates when a task involves:
- AI video or image generation (any model/platform)
- Video prompt writing or optimization
- Cinematic, commercial, or social media video production
- Product videos, brand stories, explainers, ads
- Camera movement and motion design for AI video
- Platform-specific video optimization (TikTok, YouTube, Instagram, LinkedIn)
- Higgsfield, Seedance, Kling, Sora, Veo, Wan, or any supported model
- Image generation with Soul, Nano Banana, Seedream, Flux, GPT Image
MCP Servers — Live Generation (optional)
Maxim's skill files teach Claude HOW to write generation prompts. To actually GENERATE videos/images, connect a Higgsfield MCP server:
Option A: higgsfield-mcp (Python — pip install)
pip install higgsfield-mcp
claude mcp add higgsfield -s user -e HIGGSFIELD_API_KEY=your-key -e HIGGSFIELD_SECRET=your-secret -- higgsfield-mcp
Tools: generate_image_soul, generate_video_dop, generate_speech_video, get_soul_styles, get_motions, create_character, get_job_status
Source: geopopos/geo_higgsfield_ai_mcp
Option B: higgsfield-mcp (Bun/TypeScript)
git clone https://github.com/jfikrat/higgsfield-mcp.git
cd higgsfield-mcp && bun install
claude mcp add higgsfield -s user -- bun run src/index.ts
Tools: higgsfield_generate_image, higgsfield_generate_video, higgsfield_wait_for_job, higgsfield_list_models, higgsfield_check_cost
Source: jfikrat/higgsfield-mcp
Without MCP (prompt-only mode)
If no MCP server is connected, Maxim writes the optimized prompt and the user pastes it into the Higgsfield/Seedance web UI manually. All 15 styles, MCSLA formula, camera vocabulary, and platform specs still work — just no automated generation.
API Key
Get credentials at: https://platform.higgsfield.ai
Skill Libraries
Maxim-authored Cinematic Styles (15 skills) — Maxim's original IP
Path: .claude/skills/ai-media-generation/cinematic-styles/
Coverage: 15 model-agnostic behavioral-framework-engineered cinematic styles. Each ships with a primary behavioral framework (Hook Model, AIDA, Peak-End Rule, Endowment Effect, Hero's Journey, Mirror Neurons, etc.) translated into specific cinematic structure for specific business outcomes (sales, lead gen, commitment, trust, purchase).
Entry point: cinematic-styles/README.md (moat framing + dispatch logic)
Predecessor: these styles replaced the retired community-packs/higgsfield-ai-prompts/ subtree (no-license content) in v1.0.0 with Maxim-authored equivalents that carry defensible IP. See cinematic-styles/README.md §6 for attribution notes.
External: higgsfield-ai-prompt-skill (20 sub-skills) — MIT vendored
Path: community-packs/higgsfield-ai-prompts/
Coverage: Model selection, camera, motion, audio, pipeline, prompt formula, image generation, troubleshooting, moodboards, recipes, style, and more.
License: MIT (O-Side Media). Technical reference vocabulary — no behavioral-moat framing; Maxim wraps its use via dispatch logic.
Entry point: community-packs/higgsfield-ai-prompts/SKILL.md
External: awesome-gpt-image-2 (GPT-Image-2 structured prompt library) — MIT vendored
Path: community-packs/awesome-gpt-image-2/
Coverage: 21 industrial fill-in-the-blank prompt templates + a structured atomic schema (subject/task · composition/layout · style/materials · lighting · info-hierarchy) + 500+ reverse-engineered GPT-Image-2 cases across UI · charts/infographics · posters · brand · product · character · scene · document images.
License: MIT (freestylefly). Templates + galleries stay upstream (referenced, not mirrored — see the pack's MAXIM_INTEGRATION.md). Prompt-engineering reference — no behavioral-moat framing on its own.
Entry point: community-packs/awesome-gpt-image-2/MAXIM_INTEGRATION.md
Maxim Overlay — Image Generation (GPT-Image-2) 🟢 (the moat on top)
When an image-generation task activates (triggers: image generation · GPT Image · poster · infographic · UI mockup · product image · brand image), do NOT emit a raw template. Compose through the overlay:
- Atomic schema (controllability — from the pack): fill subject/task · composition/layout · style/materials · lighting · info-hierarchy · output-spec (ratio · resolution · text legibility). Structured beats prose for batch + reuse.
- Behavioral framing (ADR-007 — cite the framework): apply the visual-behavior framework that fits the goal — pre-attentive attributes (colour/size/orientation for what must be seen first), visual hierarchy + Gestalt (grouping, single focal point), Color Psychology (palette → intended emotion/action), Fitts' Law + Nielsen heuristics for UI mockups, AIDA / Peak-End for ad creative. State which framework drove the composition.
- Brand-foundation voice (three-layer): load
.brand-foundation/personal/(+ operator + startup overlays); the palette, type, tone, and any embedded copy must pass the brand +ai-tellschecks — on-brand, not generic-AI-look. - Compliance / ethics: CSO auto-loop on regulated / person-likeness / third-party-logo content; no deepfake-class person swap without consent.
- Confidence tag (ADR-010): 🟢 only when schema + framework + brand all applied; 🔴 Maxim-UNENHANCED if a raw upstream template is used verbatim.
Generation (Layer 3 — future, gated): to produce pixels, a mxm-image MCP would call the image API (OpenAI images or an aggregator) — Maxim is model-agnostic (Maxim_MODEL_PROVIDER=openai), API key in Doppler. Not built today; this skill writes the governed prompt and the operator generates via their tool (or a connected higgsfield/image MCP).
Sub-Skill Routing
Maxim Cinematic Styles (15) — primary prompt generation
| Task Signal | Business Outcome | Style Path |
|---|---|---|
| Cinematic / film / movie scene / brand film | Commitment + emotional brand alignment | cinematic-styles/cinematic/ |
| 3D CGI / 3D animation / premium product render | Purchase confidence + premium positioning | cinematic-styles/cgi-3d/ |
| Cartoon / 2D animation / explainer | Learning retention + brand affinity | cinematic-styles/cartoon/ |
| Comic to video / motion comic / episodic | Narrative engagement + series commitment | cinematic-styles/comic-to-video/ |
| Fight / action / combat / game trailer | Arousal → sales lift | cinematic-styles/fight-scenes/ |
| Motion design / kinetic type / explainer ad | Comprehension → conversion | cinematic-styles/motion-design-ad/ |
| E-commerce / product ad / direct response | DIRECT PURCHASE (primary sales) | cinematic-styles/ecommerce-ad/ |
| Anime / Japanese animation / youth market | Tribe identity + loyalty | cinematic-styles/anime-action/ |
| Product 360 / turntable | Pre-ownership → purchase intent | cinematic-styles/product-360/ |
| Music video / performance / artist | Cultural brand equity + identity alignment | cinematic-styles/music-video/ |
| Social hook / TikTok / Reels / Shorts | LEAD GEN (top-of-funnel acquisition) | cinematic-styles/social-hook/ |
| Brand story / mission / founder film | Long-form commitment + mission alignment | cinematic-styles/brand-story/ |
| Fashion / lookbook / editorial | Aspirational desire → premium purchase | cinematic-styles/fashion-lookbook/ |
| Food & beverage / ASMR / restaurant | CRAVING → ORDER (primary sales) | cinematic-styles/food-beverage/ |
| Real estate / property / architectural | Showing request + offer (high-value lead gen) | cinematic-styles/real-estate/ |
Technical Sub-Skills (20)
| Sub-Skill | Domain | Path |
|---|---|---|
| higgsfield-camera | Camera movements & angles | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-camera/ |
| higgsfield-motion | Motion presets & dynamics | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-motion/ |
| higgsfield-audio | Audio design & sound cues | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-audio/ |
| higgsfield-pipeline | Generation pipeline & workflow | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-pipeline/ |
| higgsfield-prompt | Prompt formula & structure | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-prompt/ |
| higgsfield-image-shots | Image generation & shot design | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-image-shots/ |
| higgsfield-models | Model selection & comparison | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-models/ |
| higgsfield-cinema | Cinema techniques & language | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-cinema/ |
| higgsfield-style | Visual style & color grading | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-style/ |
| higgsfield-soul | Soul ID / Soul Cast / Soul HEX | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-soul/ |
| higgsfield-seedance | Seedance-specific features | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-seedance/ |
| higgsfield-moodboard | Moodboard creation | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-moodboard/ |
| higgsfield-recipes | Genre recipe templates | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-recipes/ |
| higgsfield-vibe-motion | Vibe-driven motion presets | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-vibe-motion/ |
| higgsfield-mixed-media | Mixed media compositing | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-mixed-media/ |
| higgsfield-apps | App integrations | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-apps/ |
| higgsfield-assist | Assistant / help utilities | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-assist/ |
| higgsfield-recall | Memory & recall system | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-recall/ |
| higgsfield-troubleshoot | Troubleshooting & debugging | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/higgsfield-troubleshoot/ |
| shared | Shared utilities & templates | community-packs/higgsfield-ai-prompts/mnt/user-data/outputs/higgsfield/skills/shared/ |
Retired in v1.0.0: Automation Skills (4)
The previously-vendored Playwright browser-automation skills (higgsfield-image-auto, seedance-auto-generate, ugc-hot-girl, ugc-video-auto) were retired in v1.0.0. Platform-specific website automation belongs to each platform's own tooling ecosystem, not to Maxim's IP portfolio. Users needing direct higgsfield.ai browser automation should install upstream tooling separately.
Maxim's role remains: generate the behaviorally-engineered prompt (model-agnostic). Delivery automation is a separate concern with different lifecycle considerations.
Behavioral Layer
Maxim Confidence Tagging
- 🟢 HIGH — Style-matched skill activated, MCSLA formula applied, platform specs validated
- 🟡 MEDIUM — General video/image prompt generated without style-specific skill
- 🔴 LOW — Raw prompt pass-through, no formula or platform optimization applied
Compliance Flags for Ad Content
- Ad disclosures: Flag when generated content requires FTC/ASA ad disclosure
- Brand safety: Validate content against brand guidelines before generation
- Platform terms: Flag content that may violate TikTok/Instagram/YouTube community guidelines
- IP/likeness: Flag when prompts reference real people, brands, or copyrighted works
Proactive Agent Loops
- Loop
brand-guardianwhen brand assets or brand voice are involved - Loop
content-strategistfor content calendar alignment - Loop
complianceskill when generating ad/commercial content - Loop
security-analystwhen generating content involving PII or likeness
Core Prompt Formula: MCSLA
All video prompts should follow the MCSLA structure:
[Model] + [Camera] + [Subject] + [Look] + [Action]
- Model: Which AI model to use (Kling 3.0, Sora 2, Veo 3.1, Wan 2.7, Seedance 2.0, etc.)
- Camera: Named camera movement preset (Dolly In, FPV Drone, 360 Orbit, etc.)
- Subject: Who/what is in frame, with specific observable details
- Look: Visual style, color grade, lighting, aspect ratio
- Action: What changes, what moves, what happens (verbs, not adjectives)
For images, use the image prompt formula:
[Shot size] + [Angle] + [Movement keyword] of [subject]. [Pose]. [Environment]. [Lighting]. [Style].
Full formula reference: references/prompt-formula.md
Platform Optimization
| Platform | Aspect | Hook Window | Duration | Notes |
|---|---|---|---|---|
| TikTok | 9:16 | 0.8s | 15-60s | Pattern interrupt in frame 1, trending audio, completion rate is king |
| YouTube Shorts | 9:16 | 2s | up to 60s | Slightly more tolerance, still mobile-first |
| Instagram Reels | 9:16 | 1.5s | 15-90s | Aesthetic-first, visual polish matters more |
| 16:9 OK | 2s | 30-120s | Professional tone, insight-driven hook, subtitles required | |
| YouTube Standard | 16:9 | 2s | any | Cinematic OK, longer storytelling, SEO title/description matter |
Full specs: references/platform-specs.md
2-Second Hook Framework
Every video prompt MUST address the first 2 seconds explicitly. Proven hook patterns:
- Extreme Close-Up Snap to Wide Reveal — disorientation then context
- Black Screen to Dramatic Light Burst — silence then explosion
- Reverse Motion — unnatural movement catches the brain
- Unexpected Scale — macro of familiar object reads as landscape
- Silent Beat Then Explosive Sound — audio contrast
- Extreme Color Shift — palette punch in 0.6s
- Fast Movement Entering Frame — motion catches peripheral vision
- Extreme Depth-of-Field Rack Focus — professional cinema technique
- Stark Geometric Contrast — visual tension
- Protagonist's Eyes Open — primal human attention grab
- Disorienting Camera Rotation — sensory engagement
- Scale Impossibility — tiny in vast or huge in small
Advanced: Stack 2-3 hooks in the opening 2 seconds for maximum impact.
Quick Reference Files
| File | Purpose |
|---|---|
references/style-guide.md |
All 15 styles — when to use each |
references/camera-vocabulary.md |
Camera movement vocabulary |
references/platform-specs.md |
Platform optimization specs |
references/prompt-formula.md |
MCSLA formula and prompt templates |
External Reference Files
| File | Purpose |
|---|---|
community-packs/higgsfield-ai-prompts/vocab.md |
Full cinematic vocabulary (camera, shot size, style, lighting, audio) |
community-packs/higgsfield-ai-prompts/model-guide.md |
Video + image model comparison tables |
community-packs/higgsfield-ai-prompts/image-models.md |
Image model specs, UI controls, pricing |
community-packs/higgsfield-ai-prompts/prompt-examples.md |
Production prompt examples by genre |
community-packs/higgsfield-ai-prompts/photodump-presets.md |
29 Photodump style presets |
Copyright (c) 2026 iSystematic Inc. Maxim product. BSL 1.1.