media-studio
Skill1 starsChanged 4 days agoMITView on GitHub
Make any image, banner, video, GIF, 3D or generative-art asset. AI image and video via Higgsfield (GPT Image 2.6, Nano Banana, Soul, Seedream, Seedance 2.5, Kling, Veo, Marketing Studio), OpenAI GPT Image or Gemini. Product photoshoots (studio, lifestyle, Pinterest pin, hero, carousel, ad pack, virtual try-on, restyle), marketplace listing images and A+ modules, UGC and unboxing ad videos, product 360 turntables and reveals, character face lock, outfit plates and character sheets (Banana Pro), cinematic scene plates, Soul ID, Virality Predictor, GLB from photos. Also blog hero, OG and social images, mockups, banners and display ads, Remotion and Hyperframes video, demo recordings, p5.js generative art, Slack GIFs, upscale and WebP. Use for: generate an image, make a video, animate this photo, remove background, inpaint, product photo, Amazon listing images, turntable, Higgsfield, Seedance, B-roll, talking head, GIF, enhance a screenshot. Logos go to brand-studio.
- Pipes a download into a shell
- Reads credentials
- Installs packages
4 other projects use a version of this.
Adopted by
Most recently taken up by
- trku052-png/social-media-youtube-skillstheir own version, first seen Nov 15, 2026
- Starlight-Intelligence/starlight-agent-skillstheir own version, first seen Nov 8, 2026
- frankxai/starlight-agent-skillstheir own version, first seen Jul 27, 2026
- nirholas/XActionstheir own version, first seen Feb 24, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- nirholas/XActions★ 658same name
- frankxai/starlight-agent-skills★ 1same name
- Starlight-Intelligence/starlight-agent-skills★ 1same name
- trku052-png/social-media-youtube-skills★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository27 other files it gives its agents.
- supabase-postgres-proSkill
- dashboard-studioSkill
- debug-labSkill
- brand-studioSkill
- ui-studioSkill
- ux-labSkill
- web-a11ySkill
- claude-builderSkill
- git-github-proSkill
- ads-and-measurementSkill
- audience-growthSkill
- conversion-labSkill
- copy-and-contentSkill
- growth-strategySkill
- lifecycle-messagingSkill
- local-seoSkill
- outbound-salesSkill
- seo-suiteSkill
- mobile-app-studioSkill
- n8n-builderSkill
- entry-optimizerSkill
- plan-and-executeSkill
- think-and-decideSkill
- research-hubSkill
- test-and-hardenSkill
- commerce-proSkill
- react-next-proSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. 0 are common, 7 are distinctive, and 8 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
7
Distinctive
This author's own choices. Worth a look.
8
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 49
Frame the job: Write down: deliverable, placement (platform and exact px from references/platform-specs.md), new or edit, one asset or a …
- Very specific: line 54
Load context before inventing anything: Read .agents/product-marketing.md or .claude/product-marketing.md if present [image, video], the …
- Distinctive: line 85
Validate like a reviewer: Check subject, style, composition, text spelled right (zoom in), edit invariants held, real-product fidelity …
- Distinctive: line 113
Gemini direct: only when the user wants Google billing or a Gemini-only feature.
- Distinctive: line 161
Use the bundled scripts as they are: no edits to them and no one-off SDK runners [imagegen].
- Very specific: line 176
Always state aspect ratio or px (on Higgsfield as the --aspect_ratio flag, not prompt text) and a style word ("clean commercial …
- Distinctive: line 187
Ask once, before any billable call or batch:
- Very specific: line 213
Run with --wait (one command, blocks, prints the URL).
- Distinctive: line 215
Every local path in a media flag is uploaded to Higgsfield.
- Very specific: line 268
Keep the original as a sibling (name-original.png).
- Very specific: line 303
Write the algorithmic philosophy (named movement, 4-6 paragraphs) as .md.
- Very specific: line 304
Read templates/viewer.html and build on it: seed section (display, prev/next, random, jump) and actions (regenerate, reset) stay.
- Distinctive: line 378
Use the API or CLI route (scripts, Higgsfield CLI, Chrome headless, Playwright MCP) before clicking through a web UI.
- Very specific: line 380
Keep keys in one credentials file outside any repo.
- Distinctive: line 413
If a member folder is missing (not installed), work from this file alone.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Media Studio
One entry point for every image, banner, video, GIF, 3D and generative-art task. It merges sixteen skills: the Higgsfield
suite (one credit pool for most image, video, audio and 3D models, plus product and marketplace enhancers), OpenAI and
Gemini image scripts, fal, two marketing-production guides (image, video), banner design, Remotion rules, a photoreal
character prompt director, a product-360 video guide, p5.js art, Slack GIFs and image enhancement. Load this file first,
pick a mode, run the unified workflow, and open a member or reference file only when a route names it. members/<name>/
points at the original skill folder; <skill> in commands means this skill's absolute folder,
~/.claude/skill-library/media/media-studio.
Pick the mode
| Task looks like | Mode | What to do / open |
| Photo, illustration, texture, sprite, concept art or scene from a prompt | AI-IMAGE | Workflow 1-9; model table in Core rules; recipes by use-case slug in members/imagegen/references/sample-prompts.md |
| Change an existing image: background swap, object removal, cut-out, relight, style transfer, composite | AI-EDIT | Same as AI-IMAGE plus the invariant rules; members/imagegen/references/prompting.md |
| Brand product visual from the user's real product photo: studio, lifestyle, Pinterest pin, hero, carousel, ad pack, try-on, surreal, seasonal restyle | PRODUCT-PHOTO | higgsfield product-photoshoot create --mode <mode>; modes and questions in references/higgsfield.md |
| Marketplace listing set: main image, secondary images, infographic, A+ modules | MARKETPLACE-CARDS | higgsfield marketplace-cards create --scope ...; marketplace rules in references/platform-specs.md |
| AI ad video: UGC, unboxing, tutorial, review, try-on, TV spot; branded ad image from a format template | AD-VIDEO | Higgsfield Marketing Studio / DTC ads in references/higgsfield.md; ad strategy from ads-and-measurement |
| Product turntable, 360 spin, reveal, beauty shot, hero product video | PRODUCT-360 | references/product-360.md, then Seedance via Higgsfield |
| Original recurring character (mascot, model, lifter): face lock, outfit plate, character sheet, outfit swap | CHARACTER | references/higgsfield-prompt-craft.md pipeline; verbatim templates in members/banana-pro-director-30/SKILL.md |
| Photoreal cinematic still or environment plate, often the start frame for a video | SCENE-PLATE | Prompt-craft reference, "Scene plates" section |
| Train a face identity for consistent images and video | SOUL-ID | members/higgsfield-soul-id/SKILL.md + photo guide; consent and paid plan first (house rules) |
| Score a finished video's hook, attention and retention | VIDEO-SCORE | Higgsfield Virality Predictor (brain_activity); A/B tests stay in conversion-lab |
| Real 3D mesh (GLB) of a product from photos | 3D-ASSET | Higgsfield multi_image_to_3d; a "3D-looking picture" is AI-IMAGE |
| Blog hero, OG preview, social post, profile or listing banner, email header | MARKETING-IMAGE | Compose in HTML/CSS, AI only for the background; sizes in references/platform-specs.md; OG and per-placement workflows in members/image/SKILL.md |
| Banner set (social cover, display ads, web hero, print) in several art directions | BANNER | members/banner-design/SKILL.md steps; 22 styles + print specs in members/banner-design/references/banner-sizes-and-styles.md |
| Show the real product UI (feature shot, store screenshots, directory gallery) | REAL-SCREENSHOT | Capture real screens (Playwright MCP or Chrome headless), frame and annotate in HTML. Never AI-generate UI [image] |
| Make an existing image sharper, larger, smaller or web-ready | ENHANCE | Local pipeline in Core rules > Enhance and optimize. Paid AI upscalers only after the user says yes |
| Templated, data-driven or batch video, motion graphics, captions | PROGRAMMATIC-VIDEO | Hyperframes vs Remotion rule below; Remotion topics in members/remotion-video-creation/rules/<topic>.md |
| Repeatable demo video of a web app flow | DEMO-RECORDING | members/video/references/product-demo-recording.md (Playwright, visible cursor, subtitles, local-app gotchas) |
| AI footage, B-roll, image-to-video, AI avatar presenter, reframe a video to 9:16 | AI-VIDEO | Higgsfield Seedance 2.5 by default (model table below); prompts in members/video/references/ai-video-prompting.md; reframe / draw_to_video workflows. Paid: ask the user |
| "Make it like this reference video" | EDIT-ANATOMY | Beat-sheet method in members/video/references/edit-anatomy.md; copy the editing grammar, never the footage |
| Generative art, flow field, particles, "art from code", seed gallery | GENERATIVE-ART | members/algorithmic-art/SKILL.md, start from members/algorithmic-art/templates/viewer.html |
| Static poster, print piece or art PNG/PDF driven by a design philosophy | POSTER | Built-in anthropic-skills:canvas-design |
| Animated GIF, Slack emoji, reaction GIF | GIF | Free first: members/slack-gif-creator/SKILL.md + its core/ helpers; paid AI GIF: gpt_image_2_6 --output_format gif |
| Voice-over, music, sound effects | AUDIO | Higgsfield seed_audio (paid, ask the user); ElevenLabs not set up; members/fal-ai-media/SKILL.md audio section for options |
| Logo, brand mark, palette, icon system | elsewhere | brand-studio (AI only for logo concepts) |
| Ad strategy, ad copy, platform ad policy | elsewhere | ads-and-measurement; this skill only makes the visual |
Unified workflow
- Frame the job. Write down: deliverable, placement (platform and exact px from
references/platform-specs.md), new or edit, one asset or a set, preview-only or project-bound, and the exact on-image text. Label every input image by role: reference, edit target, or insert/style input [imagegen decision tree + image intake]. Ask the user only for true blockers (missing copy, platform, or the real product photo); otherwise default (3 options for banners [banner-design]) and state assumptions. - Load context before inventing anything. Read
.agents/product-marketing.mdor.claude/product-marketing.mdif present [image, video], the project's brand source (docs/brand-guidelines.md, design tokens, Tailwind theme) and existing assets. Use only supplied or verified brand rules; never invent a palette, slogan or logo use [banner-design]. - Choose the cheapest route that meets the need. Climb this ladder and stop at the first rung that works
[merged from imagegen "when not to use", image approach table, banner-design visual sources]:
- existing or supplied asset (licensed);
- code-native: HTML/CSS, SVG, canvas, p5.js, Pillow; best for shapes, icons, diagrams, type-led banners and anything that must match repo assets;
- real capture: screenshots or screen recordings for anything showing the product UI; real product photos;
- AI generation or editing (paid). Use it for photos, illustrations, textures, lifestyle scenes, B-roll, product scenes built FROM the real product photo, and video. Text, logos and CTAs never live inside generated pixels when the asset is HTML-composed: they are overlaid in HTML/CSS so they stay sharp, editable and accessible [banner-design, image].
- If AI: pick the route and model with "Which generator and model" below: a Higgsfield enhancer command for product and marketplace work, a Higgsfield model id for everything else, OpenAI or Gemini direct only for the listed cases. Inside the CHARACTER pipeline, use the director's models [banana-pro-director-30]. Check current ids before promising one: they drift fast [fal-ai-media, higgsfield-generate].
- Write the prompt for the route [imagegen, higgsfield-generate, banana-pro-director-30, seedance-product-360]: direct model calls use the shared schema (40-80 words for stills, under about 200 tokens for any Higgsfield model); enhancer commands (product-photoshoot, marketplace-cards) get a short intent line, never a hand-written prompt; character and scene work follows the prompt-craft reference; video describes motion, not the start frame. Match augmentation to specificity: a detailed prompt is only normalized; a generic one gets composition, lighting and intended-use hints, never new objects, brands, slogans or palettes [imagegen].
- Gate the spend. Free checks first: OpenAI/Gemini script
--dry-run; Higgsfieldaccount status(plan, balance) andgenerate cost <id> <flags>/generate cost workflow/--cost-only. Then ask the user once per batch with route, model, count, size, quality and the estimate (credits or $), folding any open choices into labeled options and, for characters, the pre-prompt check bullets. Draft at count 1, low quality or the cheap model; scale only the winner [fal-ai-media, image cost tips, higgsfield]. Never auto-retry an uncertain billable call [imagen, higgsfield dtc-ads]. - Compose and export. Overlay copy, logo and CTA in HTML/CSS using the banner rules; capture at exact px with Chrome headless or Playwright at 2x scale after a 3-5 s font/image wait [design social-photos, banner-design]. Video overlays and sound go in post (Remotion, Hyperframes, CapCut, ffmpeg) [video, product-360 corrected].
- Validate like a reviewer. Check subject, style, composition, text spelled right (zoom in), edit invariants held, real-product fidelity (print, logo, colorway, cut), identity held across a set, safe-zone crop at real display size, contrast >= 4.5:1, file size within platform limit [imagegen, banner-design, image, banana-pro-director-30]. Iterate with one change at a time and repeat the invariants in every edit round [imagegen].
- Optimize and deliver. Download Higgsfield result URLs; web assets get the optimization checklist [image]. Save
finals into the project's asset folder by its convention, never overwrite (write
hero-v2.png), keep only selected finals, put drafts in the scratchpad. Report path (and source URL), final prompt or intent line, route/model/settings, and the real cost (balance before vs after for credits) [imagegen, higgsfield].
Video runs the same frame -> context -> route -> gate -> validate loop: script first (problem -> solution -> CTA), then pick programmatic, AI footage (Seedance), Marketing Studio ad, avatar or screen recording with the video approach table, captions always, and export per platform aspect [video]. Image-to-video from an approved still (a scene plate or a product photo) is the most controllable path [fal-ai-media, banana-pro-director-30].
Core rules and checklists
Which generator and model [new: higgsfield-generate catalog + image tree + the user's setup]
- The ladder in workflow step 3 comes first: code, screenshots and real photos beat any generator.
- Higgsfield is the default paid route for images, video, audio and 3D when
higgsfield account statusshows credits: one credit pool for most model families, a free cost estimate, product and marketplace enhancers, and the CLI is installed. The user is on the free plan, so the pool is small: check the balance before promising a batch. - OpenAI direct (
image_gen.py) when Higgsfield is out of credits or logged out, when the job needs a transparent background, a mask inpaint,input_fidelityor the script's JSONL batch, or when the user asks. Always pass--model gpt-image-2.5-sunburst(OpenAI's named replacement per the image member;-flareis the other variant) after checking OpenAI's models/deprecations page, since Higgsfield already serves GPT Image 2.6 and a newer direct id may exist. Never rely on the script defaultgpt-image-1.5(retires 2026-12-01);gpt-image-1retired 2026-10-23, so any note naming it is stale. The key is exported, so a real call bills the user's OpenAI account at once. - Gemini direct only when the user wants Google billing or a Gemini-only feature; the Nano Banana models are on Higgsfield.
- fal stays off: Higgsfield covers the same families (Nano Banana, Seedance, Kling, Veo) plus audio. Ideogram, Midjourney, Runway, HeyGen: recommend only.
| Job | Higgsfield route (id or command) | Fallback / note |
| Brand product visual from the real product photo | product-photoshoot create --mode <mode> --image <photo> | OpenAI edit with the photo |
| Marketplace listing set | marketplace-cards create --scope main|product-images|aplus|full-set | HTML composition over real photos |
| General image, design, UI mock, short text in pixels, edits | gpt_image_2_6 | OpenAI direct gpt-image-2.5-sunburst |
| Animated GIF or SVG from a prompt | gpt_image_2_6 --output_format gif|svg | GIF: Pillow builder is free |
| Cartoon or illustrated mascot, heavy texture | nano_banana_flash (Nano Banana 2) | Gemini direct |
| New original photoreal person (UGC, editorial, fashion) | text2image_soul_v2 | --soul-id for a trained identity |
| Cinematic still, concept mood | soul_cinematic | |
| Character pipeline (face lock, outfit, 3-panel sheet, outfit swap) | nano_banana_pro; GPT Image for chest-up detail; soul_cinematic for cheap face exploration | prompt-craft reference |
| One-shot face sheet or face edit from reference photos | seedream_v5_pro | |
| Environment, no people | soul_location | |
| Vector, icon, logo concept | recraft_v4_1 --model_type vector | finals in brand-studio |
| Same person or look across many images | trained Soul + text2image_soul_v2, or nano_banana_pro with reference images | Flux multi-reference (fal, off) |
| Style transfer, typography remix, a look unlike the defaults (when named) | Flux 2.0 / Flux Kontext Max | Ideogram (best text-in-image) and Midjourney: not set up |
| Cheap drafts (only when cheaper/faster is wanted) | z_image, nano_banana_2_lite | |
| Video: default (motion, image-to-video, multi-shot, 4-30 s, up to 1080p) | seedance_2_5 | |
| Video: native 4K (4-15 s) | seedance_2_0 | say why you switched |
| Video: cheaper or faster (only when asked) | kling3_0_turbo, Veo 3.1 Lite, Seedance 1.5 Pro | |
| Ad video (UGC, unboxing, tutorial, review, try-on, TV spot) | marketing_studio_video | |
| Branded ad image | marketing-studio dtc-ads generate --format-id or marketing_studio_image | |
| Reframe a video / edit it from a sketched frame | workflow reframe / draw_to_video | |
| Sound effect, ambience, voice-style, music bed | seed_audio | ElevenLabs not set up |
| GLB mesh from 1-4 product photos | multi_image_to_3d --should_texture true | |
| Score a finished video | brain_activity |
Gotchas [higgsfield-generate]: nano_banana_2 is an alias for Nano Banana PRO; Flare is its own id, never --variant;
prompt-only models (Z Image, Recraft, Soul Cast, Soul Location) reject media flags; Cinema Studio ids are missing from
model list (use model get); never pick an older model because its schema looks simpler; a model the user names stays for
follow-ups. Cost: credits vary with model, resolution, duration and count. The members only rank them: GPT Image costs more than
Banana Pro; Z Image, Nano Banana 2 Lite, Kling 3.0 Turbo, Veo 3.1 Lite and Seedance 1.5 Pro are the budget picks. Never
quote a number from memory: run the cost command. Commands, params, modes and
errors: references/higgsfield.md.
Providers on the user's machine [imagegen, imagen, fal-ai-media, image, higgsfield-*]
| Provider | Status | Best at | How to run |
| Higgsfield | CLI via Homebrew; log in once with higgsfield auth login; higgsfield account status (free) confirms | most image and video models, product/marketplace enhancers, Marketing Studio ads, Soul ID, audio, 3D, video scoring | higgsfield generate create <id> --prompt "..." --wait after generate cost; never the vendor curl | sh installer |
| OpenAI GPT Image | OPENAI_API_KEY set in the shell; openai SDK on python3.12 | transparent background, masks, edits with up to 16 input images, n=1-10, JSONL batch | python3.12 <skill>/members/imagegen/scripts/image_gen.py generate --model gpt-image-2.5-sunburst --size 1536x1024 --quality low --out <path> --dry-run (also edit, generate-batch; sizes 1024x1024, 1536x1024, 1024x1536, auto); drop --dry-run after the user's OK |
| Google Gemini image (Nano Banana) | GEMINI_API_KEY kept in a credentials file, not exported | native editing, multi-image reference | python3.12 <skill>/members/imagen/scripts/generate_image.py --model <id> --size 2K "<prompt>" <new>.png (1K, 2K, 4K), key exported for that one command only; helper uses generateContent |
| fal.ai | not set up: no key, no MCP | same families as Higgsfield | Off; recommend only if Higgsfield lacks a model the user needs |
| Ideogram, Recraft direct, Runway, HeyGen, ElevenLabs | not set up | Ideogram text-in-image, Runway motion control, HeyGen avatars (has MCP), ElevenLabs voice | Recommend with the install/signup step; the user decides |
Use the bundled scripts as they are: no edits to them and no one-off SDK runners [imagegen].
Prompt schema (direct model calls) [imagegen + image]
Use case: <slug: photorealistic-natural | product-mockup | ui-mockup | infographic-diagram | logo-brand |
illustration-story | stylized-concept | historical-scene | text-localization | identity-preserve |
precise-object-edit | lighting-weather | background-extraction | style-transfer | compositing | sketch-to-render>
Asset type: <where it will be used, px>
Primary request: <one line>
Input images: <Image 1: role; Image 2: role>
Scene/backdrop | Subject | Style/medium | Composition/framing | Lighting/mood | Color palette | Materials/textures
Text (verbatim): "<exact text>"
Constraints: <change only X; keep Y unchanged> Avoid: <no logos, no watermark, no extra text>- Order: scene -> subject -> details -> constraints. Name the intended use; it sets polish level [imagegen].
- Always state aspect ratio or px (on Higgsfield as the
--aspect_ratioflag, not prompt text) and a style word ("clean commercial photography", "flat illustration", "3D render"). - Photoreal: camera language (lens, depth of field, light direction). Leave negative space where copy will sit.
- Text in pixels: 3-5 words max, quoted verbatim, tricky words spelled letter by letter; anything longer goes in an overlay [imagegen, image]. Logo placement by prompt is unreliable: add logos in post [image].
- Multi-image input: refer to images by index and say what each contributes. Reuse a
seedwhen iterating [fal-ai-media]. - Image-to-image: describe the change, not the input; phrase positively ("tack sharp", not "no blur") [higgsfield-generate].
- For batches: test 3-4 style prompts, pick the winner, save that exact prompt as the template, feed the winner back as a style reference, then vary only the subject [image batch workflow].
Paid-call request and delivery report [imagegen report + fal cost check + imagen data rules + higgsfield cost]
Ask once, before any billable call or batch:
Paid generation request: <route> <model or command/mode>, <n> output(s) at <size/res/duration>, quality <...>
Estimated cost: <N credits from `generate cost` (balance B) | $ from provider pricing | unknown until a count-1 run>
Uploads: <files that go to the provider, or none>. Prompt has no private data.
Why AI and not code/screenshot/photo: <one line>. Draft first at <cheap setting>, final only for the pick.
Options (defaults marked): <labeled choices, if any>. OK?Deliver with: final file path(s), px and source URL, the final prompt or intent line, route/model/settings, real cost, what was checked (text, invariants, product fidelity, crop), and any assumption made.
Never AI-generate these [image, video]
Product UI (hallucinated interfaces), real locations people know, logos as finals, paragraphs of text, and the product being sold from text alone. A product image or video is AI-made only when conditioned on the user's real product photo or design file (PRODUCT-PHOTO, MARKETPLACE-CARDS, PRODUCT-360, AI-EDIT composite).
Real product in AI imagery [image + higgsfield product members + seedance-product-360]
- Pass the real product photo or the design file every time (
--image); text-only product renders never ship. - Compare each output with the real product: print artwork position, scale, colors and spelling; logo; colorway; garment cut and fit; proportions. Prints and small text are where models drift most: reject drifted ones rather than fixing in prose.
- Marketplace images: check the target marketplace's current rules before upload (Amazon main image: pure white, product
about 85% of the frame); a
full-setis 13 jobs, so cost it as 13. - Label AI-made content where TikTok, Meta or Amazon require it.
Higgsfield job rules [higgsfield-generate, product-photoshoot, marketplace-cards, soul-id]
- Run with
--wait(one command, blocks, prints the URL);--jsononly for chaining. Timeouts may still bill: checkhiggsfield generate list --jsonbefore any rerun. - Every local path in a media flag is uploaded to Higgsfield. Product/brand-kit import from a URL: the user's own store or site only.
- Enhancer commands own their prompt: pass a short intent; never call
gpt_image_2directly for product shots (worse output) and never paste the enhanced prompt back. - Marketing Studio: an ad reference video OR hook/setting blocks, never both;
--hook_id/--setting_idonly onugc,ugc_how_to,ugc_unboxing,product_review,ugc_virtual_try_on; arrays as@file.json; DTC ads need a format the user picks; custom avatars only from consented faces. - Discovery: run the unfiltered
higgsfield model list --json/workflow listbefore saying a model or feature is missing.
Character and scene prompt craft [banana-pro-director-30; full guide in references/higgsfield-prompt-craft.md]
- Order: existing-or-new check -> text spec -> Mode 0 face lock -> Mode 1 base outfit plate -> Mode 2A 3-panel sheet (headless front, rear, tight chest-up face) -> Mode 3 scene plates on request; Mode 4 detail and Mode 5 outfit swap on explicit request. Never skip or combine steps: every later image anchors to the locked face.
- Reference plates (Modes 0, 1, 2, 5) close with the flat grade: one even 18% gray value, shadowless matched-fill light, zero cast shadow, true skin and fabric color, baseline black camisole or ribbed tank for face locks. Gray, because video models amplify high-contrast edges; flat, because baked shadows get inherited by every later generation.
- References carry identity and geometry: one short visual handle per subject, words spent on composition, pose, light and the close; under about 2,500 characters with strong references.
- Scene plates: five unlabeled DP-prose paragraphs, resolution-aware detail (drop what the lens couldn't resolve), positional words instead of coordinates, the cinema-stack close with "photographed not generated".
- Universal: no character names or ages, flattering realism (fine even skin, never blemished), closed-lip default expression, no negative-prompt blocks, aspect ratio only as a flag.
Product 360 and reveal video [seedance-product-360, corrected; full guide in references/product-360.md]
- Beats: 2 s hook matched to the category's energy -> hero angle -> rotation (slow 10-15 s, medium 5-8 s, fast 2-4 s per 360) -> 2-3 macro details (2-5 s total) -> optional context -> finale on the hero angle.
- Product fills 50-80% of the frame; product-to-background contrast at least 3:1; light per category and material table.
- Seedance limits: 2.5 = 4-30 s up to 1080p; 2.0 = 4-15 s up to 4K. Longer stories are several shots (start/end frames from the same shoot) joined in the edit. One shot per prompt, under about 200 tokens; the long template is a planning sheet.
- Names, specs, prices and CTA are overlays in post; sound is native Seedance 2.0 audio,
seed_audio, or music in the edit.
Layout rules for banners, social and OG images [banner-design + design social-photos]
- Safe zone: critical content in the central 70-80%; nothing within 50-100 px of edges; YouTube channel art keeps content inside 1546 x 423 of 2560 x 1440; Meta/Instagram central 80% (UI chrome).
- 3-zone hierarchy: top = logo or value prop, middle = message + visual, bottom = CTA.
- One CTA, high contrast, bottom-right, >= 44 px tall, action verb (Get, Start, Claim, Shop).
- Max 2 typefaces; body >= 16 px, headline >= 32 px; contrast >= 4.5:1; ads <= 7 words a line and 3 lines.
- Text share: paid social ads under 20% of the canvas; social covers about 60/40 image to text.
- Print: 300 DPI (150 for large format), CMYK, 3-5 mm bleed, headlines 70 pt+ for 3-5 m viewing.
- File names: project convention first; otherwise
assets/banners/<campaign>/<style>-<w>x<h>.png, date prefixYYMMDD-for time-bound campaigns [banner-design].
Export (free, local) [design social-photos, banner-design]
CHROME="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"
"$CHROME" --headless --disable-gpu --hide-scrollbars --force-device-scale-factor=2 \
--window-size=1200,630 --virtual-time-budget=5000 --screenshot=og.png "file:///abs/path/og.html"Then confirm pixel size (sips -g pixelWidth -g pixelHeight og.png) and that fonts loaded. With 2x scale the file is
2400 x 1260: downscale to the platform size if it has a hard limit. Playwright MCP (browser_resize +
browser_take_screenshot) is the alternative when the page needs interaction first.
Enhance and optimize [image + image-enhancer, made concrete]
- Measure first: dimensions, format, file size, what is wrong (blur, noise, artifacts, too large).
- Keep the original as a sibling (
name-original.png); writename-enhanced.png. Tune for the intended use (web, slides, print). Folders: loop the same steps per file and report size and dimensions before vs after. - Screenshots: re-capture at 2x device scale instead of upscaling; upscaling adds no real detail.
- Photos/graphics (Pillow is on
python3.12): upscale at most 2x withImage.LANCZOS, thenImageFilter.UnsharpMask(radius=2, percent=80-150, threshold=3);MedianFilter(3)only on visibly noisy photos. - Format: WebP default (photos and graphics), AVIF for max compression, PNG for transparency and UI text, SVG for logos
and icons, JPEG as fallback.
cwebp -q 80 in.png -o out.webpis installed;sips -Z <px>resizes. - Web checklist: quality 75-85 for photos, near-lossless for screenshots; hero under 200 KB; resize to display size;
explicit
width/height(no layout shift);loading="lazy"below the fold; descriptive alt text; CDN auto-format. - OG/social preview: 1200 x 630 plus
og:image,og:image:width,og:image:height,twitter:card=summary_large_image,twitter:image; for many pages generate per-page OG images with@vercel/og/ Satori [image].
Video [video + remotion-video-creation + higgsfield-generate]
- Prompt = subject + action + camera move + style + mood; name the camera move (dolly, pan, static, pull back). Image-to-video is more controllable than text-to-video [fal-ai-media]. No readable text from AI video: overlay it.
- Seedance 2.5 modes:
t2v(prompt only, no media),omni_reference(any references incl.--start-image/--end-image),video_edit,video_extension. Reframe an existing clip with thereframeworkflow instead of regenerating it [higgsfield-generate]. - Captions always (most social video plays muted). Aspect: 9:16 social, 16:9 YouTube/site, 1:1 or 4:5 feed.
- Hyperframes vs Remotion: plain HTML/CSS + GSAP, short templated clips, no React project -> Hyperframes (one HTML file
per composition: root
data-composition-id,data-duration,data-width,data-height; clipsclass="clip"withdata-start,data-duration,data-track-index; paused GSAP timeline registered inwindow.__timelines). React project, spring physics, rich captions, 3D/Lottie, or Lambda batch rendering -> Remotion. - Remotion hard rules: every animation comes from
useCurrentFrame(); write timings in seconds timesfpsfromuseVideoConfig(); CSS transitions/animations and Tailwindanimate-*classes do not render; referencepublic/files withstaticFile(); premount every<Sequence>; composition props must be JSON-serializable. Open the matchingrules/*.md(captions, audio, fonts, transitions, charts, 3d, measuring-text) before writing that part. - Pin tool versions in commands (
npx hyperframes@<version>,npx create-video@<version>) and say in one line what gets installed. Remotion's license is free only below a company-size limit: check remotion.dev/license before commercial use. - Avatars: Marketing Studio UGC presenters (Higgsfield, set up), HeyGen (best lip-sync, MCP, not set up) or Synthesia (corporate). Not for founder-authentic content or UI walkthroughs.
- Repurposing chain: clean up (Descript) -> auto-clip (Opus Clip) -> captions/styling (CapCut) -> schedule. Virality Predictor can score the cut's hook before posting; it is an attention proxy, not a sales result.
Generative art [algorithmic-art]
- Write the algorithmic philosophy (named movement, 4-6 paragraphs) as
.md; weave in one quiet conceptual reference. - Read
templates/viewer.htmland build on it: seed section (display, prev/next, random, jump) and actions (regenerate, reset) stay; algorithm, params object and parameter controls change. - Seeded everything (
randomSeed(seed); noiseSeed(seed)); same seed must give identical output; params are quantities, scales, probabilities, ratios, angles, thresholds. Thoughtful palettes, real-time performance. - Single self-contained HTML; p5.js from
cdnjs.cloudflare.com/ajax/libs/p5.js/1.7.0/p5.min.js(allowed by the Artifact page policy). - Published as an Artifact, the page cannot start downloads: drop the Download PNG button or use the artifact file
hand-off capability (
artifact-capabilitiesskill). In a local HTML file the button works.
GIFs [slack-gif-creator]
- Slack: emoji 128 x 128 (under 3 s), message GIF 480 x 480; 10-30 fps; 48-128 colors.
- Build frames with Pillow primitives through
core.gif_builder.GIFBuilder; ease motion withcore.easing.interpolate(ease_out, bounce_out, back_out, elastic_out); check withcore.validators.is_slack_ready. - Make it look crafted: lines width >= 2, gradient backgrounds, layered shapes, glows, contrasting outlines; no emoji fonts.
- Motion recipes in the member: shake, pulse/heartbeat, bounce, spin/wobble, fade, slide with overshoot, zoom, particle burst; combine two for richer loops.
- Optimize only when asked: fewer frames, fewer colors, smaller size,
remove_duplicates=True,optimize_for_emoji=True. - Setup: Pillow is on
python3.12;imageioandnumpyare not, so install them pinned into a scratchpad venv and say so. Import the helpers withsys.path.insert(0, "<abs path of members/slack-gif-creator>"); run from the scratchpad. - Photoreal or illustrated animated GIFs the builder can't draw:
gpt_image_2_6 --output_format gif(paid), then validate size and colors with the same validators.
When members disagree
| Between | Conflict | Rule |
| imagegen vs Claude Code reality | imagegen prefers a built-in image_gen tool and allows its CLI only "when explicitly asked" | Claude Code has no built-in image tool, so the CLI is the OpenAI path. "Explicitly asked" becomes "the user approved this paid call" |
| imagegen script vs OpenAI retirements | script defaults to gpt-image-1.5 (retires 2026-12-01); fal-ai-media's note still offers gpt-image-1 (retired 2026-10-23) | Always pass --model gpt-image-2.5-sunburst after checking OpenAI's page; treat any gpt-image-1 mention as stale |
| higgsfield-generate vs the user's spend rule | vendor: don't pre-estimate cost or offer cheaper models unless asked | Keep the quality default model, but always run the free cost check and ask once per batch; offer the cheap draft as an option, never swap silently |
| higgsfield-generate vs product-photoshoot vs banana-pro vs the user's autonomy rule | one question at a time / up to 4 labeled questions / a pre-prompt check every prompt / ask only blockers | One message: defaults as labeled options + pre-prompt bullets (references first) + the cost line. Skip it for small tweaks on an approved prompt |
| product members vs "never AI-generate the product" | product-photoshoot, marketplace-cards and product-360 generate the product | Allowed only when conditioned on the real product photo; verify against the real product; text-only product renders never ship |
| higgsfield-generate auto-pick vs banana-pro-director-30 | vendor: character sheets -> Seedream 5.0 Pro, Nano Banana Pro only when named; director: Banana Pro for face lock and sheets | Inside the director pipeline its models count as named (its prompts are written for them); one-shot face sheets from photos -> seedream_v5_pro |
| banana-pro-director-30 internal | "cinema stack closes Modes 0-5" vs "flat grade, no exceptions"; Mode 4 beauty lighting; 6-panel three-point light; Mode 0 "white background" | Every reference plate gets the flat 18% gray grade; the cinema stack only on finished stills (Mode 3 prose, Mode 4 as a deliverable, white-card stills) |
| banana-pro-director-30 internal | "no @image tags" vs the locked Mode 5 prompt using @image1/@image2 | Mode 5 is the one exception; pass the outfit reference first and prove the order on a count-1 run |
| banana-pro-director-30 vs higgsfield prompt safety | real brand names allowed vs avoid trademarks (ip_detected) | the user's own brands fine; third-party names only when the user asks; otherwise use original designs |
| seedance-product-360 vs media-studio video rules | text overlays and 20-28 s single prompts inside Seedance; Seedance "5-60 s" | Overlays in post; one shot per prompt under ~200 tokens; respect 4-30 s (2.5) and 4-15 s (2.0) |
| seedance-product-360 vs evidence | ray tracing, particle systems, text animation, "Higgsfield logo fine on Amazon" | Dropped as unsourced; platform rules checked on the platform |
| higgsfield delivery vs imagegen delivery | URLs only, no ids vs local path, prompt, provider, cost | Both: URL + downloaded local file + model + credits spent; no enhanced prompt unless asked |
| higgsfield bootstrap vs the user's security rules | run the vendor curl-pipe-sh installer when the CLI is missing | install the CLI from Homebrew; never run the installer; upgrades only with the user's OK |
| product-photoshoot vs prompt schema | "never write the GPT Image 2 prompt yourself" vs the shared schema | Schema for direct model calls; enhancer commands take a short intent line |
| banner-design vs imagegen | banner: generated visuals carry no text; imagegen: quote exact text in the prompt | HTML-composed assets: no text in pixels. Flat raster pieces where text must be baked in (poster, sticker): 3-5 words, verbatim, use a text-strong model, verify spelling |
| image vs brand-studio's design member vs Recraft on Higgsfield | image: AI logos are poor; design: generate logos with Gemini in 55 styles | AI makes logo concepts only (paid, ask); finals are hand-built SVG in brand-studio |
| imagegen vs image | imagegen: do not add details the user did not imply; image: always add style, lighting, aspect | Always add format and style words; add creative content (objects, brands, palettes) only if implied |
| algorithmic-art vs the user's brands | keep Anthropic colors and fonts in the viewer | Keep the template's structure and controls; for the user's brand work, swap the UI colors/fonts for the project's tokens. Keep Anthropic styling for neutral exploration |
| video vs remotion-video-creation | Hyperframes "recommended for agents" vs Remotion rules | Choose by the rule in Core rules > Video; don't port an existing Remotion project to Hyperframes |
| banner-design vs the user's autonomy rule | gather 6 answers via questions first | Ask only blockers; default the rest and list assumptions in the delivery |
| imagegen vs banner-design vs brand | output paths output/imagegen/ vs assets/banners/<campaign>/ vs brand asset naming | Project convention first; then the asset-type default; drafts in the scratchpad |
| fal-ai-media vs reality | "unified media via fal MCP" | fal is not set up and Higgsfield covers its models; recommend only if a model is missing |
| image-enhancer vs honesty | promises "upscaled to 4K, enhanced clarity" | Upscaling adds no detail; re-capture when possible and report what was actually done |
| code-drawn site imagery vs AI photos/video | banana-pro and seedance notes favour code-drawn scenes; Higgsfield allows AI imagery | Follow the user's brand: code-drawn scenes can stay the signature, AI hero loops, lifestyle, try-on and 360 are allowed, each batch cost-checked |
Combine with
- Built-in skills:
anthropic-skills:canvas-design(posters, static art),anthropic-skills:algorithmic-art(same as the member),artifact-designbefore publishing any page,artifact-capabilitiesfor file hand-off or stored picks,frontend-design:frontend-designfor type-led banners,datavizwhen an image carries a chart. - Superpowers:
superpowers:brainstormingbefore a new visual direction, campaign or recurring character;superpowers:dispatching-parallel-agentsfor large variant sets (one agent per size or style; paid jobs still go through one approval);superpowers:verification-before-completionbefore "done". - Agent prompts (
~/.claude/skill-library/agents/voltagent/):06-developer-experience/visual-asset-generator.md(asset checklist for app icons, favicons, OG, social banners; skip itsnpm install -g prompt-to-assetstep, that package is unreviewed),01-core-development/ui-designer.md(visual polish of composed assets),01-core-development/frontend-developer.md(Remotion or Hyperframes builds),08-business-product/content-marketer.md(what the asset should say and where it goes),05-data-ai/prompt-engineer.md(tighten a long prompt template under the ~200-token limit). - Other super-skills: brand-studio (logo, palette, tokens), ads-and-measurement (ad creative strategy and ad specs), copy-and-content (headlines on the asset), conversion-lab (hero image placement, A/B tests of creatives), commerce-pro (uploading product media and listings to Shopify or marketplaces).
House rules (the user)
- Paid generators ask first: Higgsfield credits (every job, incl. Virality Predictor and Soul training), OpenAI images,
Gemini, fal, ElevenLabs, HeyGen, Runway and any AI upscaler. One question per batch with route, model, count, settings
and the estimate; the dry run,
account statusandgenerate costare free. - Use the API or CLI route (scripts, Higgsfield CLI, Chrome headless, Playwright MCP) before clicking through a web UI; if a browser is truly needed, say why in one line.
- Keep keys in one credentials file outside any repo; export
GEMINI_API_KEYonly for the one command that needs it; never print or paste a key and never open~/.config/higgsfield/. - Uploads to Higgsfield (product photos, design files, faces, videos, product or brand URLs) only for files the user approved for that batch; URL imports only from the user's own store or site.
- Faces: Soul ID, face locks, custom avatars and outfit swaps use only the user's own face with their say-so or a person's clear
written OK; never recreate a real person's likeness. Soul training needs a paid Higgsfield plan (check the plan with
higgsfield account status). - Anything that publishes, posts or uploads an asset to a live store, marketplace, site or social account asks the user first; label AI-made content where TikTok, Meta or Amazon require it and check marketplace image rules before a listing upload.
- Store and conversion work: every visual must help a visitor buy. Real product imagery near the buy button, no explainer graphics; cost-check each batch.
- Installing a new tool (fal MCP, Hyperframes, imageio, a Higgsfield upgrade) means a reviewed source, a pinned version and one line telling the user; no curl-pipe-bash installers. Members without an upstream license are for personal use only; never republish their text.
Members
| Member | Best at | Path |
| imagegen | OpenAI GPT Image CLI (generate/edit/batch, masks, transparency), use-case taxonomy, prompt schema, invariant-first editing, save-path hygiene | members/imagegen/SKILL.md, references/{prompting,sample-prompts,cli,image-api}.md, scripts/image_gen.py |
| imagen | Gemini image helper (stdlib only, PNG validation, no auto-retry of billable calls) | members/imagen/SKILL.md, scripts/generate_image.py |
| fal-ai-media | One API for image, video and audio models; seeds; cost estimation; model discovery | members/fal-ai-media/SKILL.md |
| image | Marketing image production: model-choice tree, placement workflows (hero, social, mockups, banners), web optimization, OG tags, OpenAI retirement dates | members/image/SKILL.md, members/image/references/ai-image-prompting.md |
| banner-design | Multi-option banner process, 22 art directions, safe zones, CTA and type rules, print specs, export naming | members/banner-design/SKILL.md, members/banner-design/references/banner-sizes-and-styles.md |
| video | Video approach choice (programmatic, AI, avatar, repurpose), AI video model table, edit-anatomy beat sheet, scripted demo recording | members/video/SKILL.md, references/{ai-video-prompting,edit-anatomy,product-demo-recording}.md |
| remotion-video-creation | 28 Remotion rule files (animation, sequencing, captions, audio, fonts, transitions, 3D, charts) | members/remotion-video-creation/SKILL.md, rules/*.md |
| algorithmic-art | Philosophy-first p5.js generative art with seeded variation and an interactive viewer template | members/algorithmic-art/SKILL.md, templates/viewer.html |
| slack-gif-creator | Slack GIF limits, GIFBuilder, easing, validators, animation recipes | members/slack-gif-creator/SKILL.md, core/*.py |
| image-enhancer | Enhancement intent (screenshots for docs, social, slides; keep originals); the concrete method is in Core rules | members/image-enhancer/SKILL.md |
| higgsfield-generate | Higgsfield CLI: model catalog and auto-picks, media flags, cost commands, Marketing Studio ads, DTC ads, brand kits, workflows, Virality Predictor, 3D, audio, errors | members/higgsfield-generate/SKILL.md, references/{model-catalog,media-inputs,workflows,marketing-*,prompt-engineering,troubleshooting}.md; merged in references/higgsfield.md |
| higgsfield-product-photoshoot | 10 product-photo modes with a backend prompt enhancer, mode tie-breakers, short labeled interview | members/higgsfield-product-photoshoot/SKILL.md |
| higgsfield-marketplace-cards | Marketplace main, secondary and A+ image sets by scope or asset list | members/higgsfield-marketplace-cards/SKILL.md |
| higgsfield-soul-id | Face identity training: photo rules, variants, use with Soul models | members/higgsfield-soul-id/SKILL.md, references/{photo-guide,troubleshooting}.md |
| banana-pro-director-30 | Photoreal character pipeline (face lock, outfit plates, 3-panel sheets, outfit swap), flat-grade reference plates, cinema-prose scene plates, anti-AI-look render rules | members/banana-pro-director-30/SKILL.md; merged in references/higgsfield-prompt-craft.md |
| seedance-product-360 | Product video craft: hooks, hero angles, light per category and material, rotation and camera library, multi-angle photo plans, overlay timing, review checklist | members/seedance-product-360/SKILL.md; corrected and merged in references/product-360.md |
If a member folder is missing (not installed), work from this file alone.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

