Creative Media

creative-media — a repeatable produce→review→publish pipeline

What this sub-skill is for

Standing up a creative-media project: a repeatable pipeline that turns a concept into published assets — AI image/video/audio generation, a YouTube or short-form channel, an ads operation, thumbnails, or brand visuals. Loaded by new-project for any "make videos / generate images / run a channel / produce ads / design thumbnails / brand visuals" request. The whole skill is organized around one idea: build a pipeline (concept → generate → review → publish) that can run again, not just one lucky asset.

Mandatory grill-questions (fold into the Definition of Ready)

Lock these before generating anything:

• Medium — image, video, audio, or mixed? Decides the generation model + the whole pipeline.
• Platform + format/aspect — YouTube long-form (16:9), Shorts/Reels/TikTok (9:16), feed posts

(1:1/4:5), display ads (set sizes)? Pin the exact aspect ratio(s) and duration now.

• Brand / style — palette, typography, tone, references. Harvest tokens from references with

design-harvest. Consistency across assets matters more than any single hero piece.

• Volume / cadence — how many assets per week, and the publishing rhythm? The pipeline must

sustain that cadence, not just produce one piece.

• Identity consistency — recurring faces (use higgsfield-soul-id to train a Soul) or recurring

products (use higgsfield-product-photoshoot)? Lock identity before batch-generating.

• Edit / assembly flow — where raw generations get cut, captioned, scored, and exported (an NLE,

a script, higgsfield's tools). No emojis in any on-screen text/overlays.

• Publish + rights — who posts, where, and is every input (faces, music, product, stock)

cleared for the intended use? AI-disclosure required on the target platform?

• Success metric — what "this works" means: a published asset, a tested hook, a virality_predictor

score over a bar, or a filled content calendar.

Project sub-agents to generate (into /.claude/agents/)

• concept-writer (delegate-by-default) — turns the brief into concepts, hooks, scripts, and

generation prompts on-brand and on-format; keeps a backlog feeding the pipeline's cadence.

• asset-generator (delegate-by-default) — runs the higgsfield generation: generate_image,

generate_video, generate_audio, applying the trained Soul (--soul-id) or product-photoshoot mode for consistency, and organizing outputs by campaign/episode.

• thumbnail-designer — designs thumbnails/covers/key art that read at small sizes, on-brand,

emoji-free; chains design-harvest for reference patterns and the higgsfield image tools.

• virality-checker (delegate-by-default for video) — runs virality_predictor

(brain_activity) on cut videos for hook strength, attention, retention, and distraction risk; reports the score honestly and flags weak hooks before publish.

• Plus the kit default reviewer (brand/format/rights review before publish).

Tools / CLIs / MCP / skills needed

Check in environment-readiness; offer install, never auto-install:

• CHAIN the higgsfield MCP + skills as the generation core: higgsfield-generate

(generate_image / generate_video / generate_audio, image-to-image, image-to-video, Marketing Studio for ads with avatars/products/hooks), higgsfield-soul-id (train a face for identity-consistent output), higgsfield-product-photoshoot (brand/product visuals), higgsfield-marketplace-cards (listing/product image sets), and virality_predictor for video. Use media_upload_widget / media_import_url for the user's own input media.

• CHAIN design-harvest — pull palette/typography/spacing/component tokens from reference sites

for the brand system and thumbnails. CHAIN framer-inspiration if the project ships a site.

• CHAIN the pptx skill for pitch/brand decks and the magic MCP for any web UI/overlay

templates. docx for scripts/treatments if wanted.

• Editing/encoding — ffmpeg (cut/caption/encode/aspect-convert), an NLE if hand-edited.

Verify ffmpeg is present. CHAIN global skills: deep-research (platform/format/trend facts), verify, code-review (if the pipeline is scripted), cli-anything (wrap a GUI editor like Shotcut/Inkscape headlessly).

File / asset nudges (on top of the base set)

Beyond CLAUDE.md, PROJEKT_.md, TASKS.md, DONE.md, README, .claude/:

• BRAND.md — palette, type, tone, do/don'ts, the trained Soul/product reference ids, and the

emoji-free overlay rule. The consistency contract.

• PIPELINE.md — the repeatable concept → generate → review → publish steps, who/what runs each,

and the per-format spec (aspect, duration, safe areas).

• CALENDAR.md — the publishing cadence + a content backlog.
• assets/ split into raw/ (generations), edited/, published/, thumbnails/; large media

.gitignored or Git-LFS.

• assets/RIGHTS.md — provenance + license + usage-right + AI-disclosure status per input/output.

Non-optional.

Stack defaults & done-bar

Default stack: higgsfield for generation (GPT Image 2-class for images/design/text, Seedance for video, generate_audio for sound; Soul for consistent faces, product-photoshoot for products), design-harvest for brand tokens, ffmpeg for assembly/encode/aspect, virality_predictor to score video, pptx for decks. Assets organized in assets/, big media kept out of git or in LFS. Done-bar (all true): a repeatable produce→review→publish pipeline is documented in PIPELINE.md and runnable again at the chosen cadence; the first real asset(s) are generated, on-brand, in the correct format/aspect, and emoji-free; identity consistency holds where required (Soul/product); for video, hooks pass the virality_predictor bar or weak ones are flagged; every input/output has a recorded right in assets/RIGHTS.md.

Guardrails

• Disclose AI-generated media where the platform or law requires it; record disclosure status in

RIGHTS.md. Don't pass synthetic media off as a real photograph where that would mislead.

• Rights / licensing — every face, voice, song, product, and stock input must be cleared for the

intended commercial use before it ships; record provenance per asset.

• No deepfakes of real people without consent — only generate a real person's likeness/voice

with their explicit permission (the Soul-ID identity flow is for consented faces only).

• Honest performance — report virality_predictor scores straight; don't dress up a weak hook

as strong, and don't claim reach/engagement you haven't measured.

• No emojis in any on-screen text, overlay, thumbnail, caption, or UI (user's standing rule) —

typographic symbols or designed glyphs only.

• Never commit secrets / platform API keys / publish tokens.
• Commits under the user's own name only (Skryx-L-A); never add Claude as a co-author.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Skryx-L-A
• Source: Skryx-L-A/project-kit
• License: MIT

Install and usage instructions live in the source repository linked above.