Turn Claude Code + your terminal into a creative studio — design, video, audio, transcription, copywriting, frame-aware video study in one install.
| Link | Section | What it does | Time |
|---|---|---|---|
| What this is | Overview | TL;DR — what the install actually does | ~1 min |
| Why this exists | Context | Why AI ships ugly UI by default and how this fixes it | ~2 min |
| Cheat sheet | Reference | Every slash command / skill / MCP call you'll use | ~2 min |
| Install | Setup | One-liner (curl) or clone + run | ~1 min |
| Manual steps | Setup | The 5 tools that want an OAuth click or API key | ~1 min |
| What gets installed | Reference | Table of every skill / MCP / binary | ~1 min |
| How I use each piece | Context | The reason each tool is in the box | ~3 min |
| Taste skills | Reference | 8 variants, 3 knobs, when to pick which | ~2 min |
| Update | Maintenance | Pull the latest without reinstalling | — |
| Uninstall | Maintenance | Reverse everything, prompt before touching ffmpeg |
— |
| The maxxing series | Meta | Sibling repos: cli-maxxing + task-maxxing | — |
| License | Meta | MIT | — |
creativity-maxxing is a single-install creative studio for Claude Code. You run one script and your terminal picks up:
- Taste-enforcing design skills that stop Claude from shipping the usual centered-hero + three-card-grid slop.
- A natural-language component generator (21st.dev Magic) that pulls from a pro library instead of hallucinating JSX.
- Canva wired into your terminal — design work, brand kits, exports, conversational edits.
- Remotion so Claude writes React-native video compositions instead of re-inventing timing code every time.
- A full transcription stack — drop any YouTube / Instagram / TikTok link and get the transcript back, locally, no API key.
- Higgsfield / Seedance 2.0 prompt skills — 15 video-style prompt engineers pre-wired for Claude.
/copywriting— a master-trained anti-AI-slop copy filter. Bernbach, Hegarty, Abbott, Trott, Wieden, Sugarman, Sackheim, Schwartz, Bencivenga, Gossage, Krone, McElligott. Auto-activates on headlines, hero copy, body copy, CTAs, manifestos, proposal docs, landing pages, ad copy, brand voice work, naming, and "rewrite this paragraph."/watch— drop any video URL (single video, full YouTube channel, or playlist) and get back timestamped, frame-aware study notes. Splits video into scene-change frames + a transcript (captions free, local whisper.cpp fallback — runs key-free off the media-module stack), then Claude reads every frame as an image and writes a section-per-scenenotes.mdto~/watch/library/. Channel mode rolls up anindex.mdwith cross-channel synthesis (recurring hooks, formats, cloneable moves).
It's built for Claude Code in the terminal, not the desktop app.
AI is genuinely bad at design out of the box. Every model defaults to the same centered hero, same three cards, same Lucide icon trio, same gradient. That's what taste exists to block, and what 21st.dev Magic + UI/UX Pro Max exist to replace with actual pro references — some free, some paid, all better than the default output.
The other half is the thing I could never make work before: I hate every online transcription tool. They paywall the good ones, the free ones mangle timestamps, and nothing lets me drop a link and walk away. So this install wires up yt-dlp + local Whisper so you paste any link — a YouTube video, an Instagram Reel, a TikTok, a random tweet with audio — and Claude transcribes it locally. No API key, no upload, no paywall.
Add Canva into the terminal and the whole loop gets weird in a good way: you can design conversationally. "Make me a launch banner, three variants, green-leaning, thin slab font" — and it just happens. No tab-switching, no mouse, no starting from a blank template.
There's more in here (Remotion for programmatic video, the Higgsfield/Seedance prompt pack for short-form, the full taste pack), but those three — the taste filter, the transcription pipeline, and Canva-in-terminal — are the unlocks.
Important
A few of these need one extra step after the install finishes — typically an OAuth click in your browser or pasting a free API key. The install script tells you exactly which ones and gives you the command / link. See Manual steps.
One-liner:
bash <(curl -fsSL https://raw.githubusercontent.com/fidgetcoding/creativity-maxxing/main/install.sh)Or clone and run locally:
git clone https://github.com/fidgetcoding/creativity-maxxing.git
cd creativity-maxxing
bash install.shWant Gamma too? It's opt-in (it fails to connect without an API key from gamma.app/api):
bash <(curl -fsSL https://raw.githubusercontent.com/fidgetcoding/creativity-maxxing/main/install.sh) --with-gammaThe installer runs design/install.sh, then media/install.sh, then copywriting/install.sh, then watch/install.sh in order, refuses to start if claude or ~/.claude/skills/ is missing, and is idempotent — re-run any time without duplicating installs. Idempotency marker: ~/.claude/.creativity-maxxing-installed (delete to force a full reinstall).
Flags:
--with-gamma— also register the Gamma MCP (default off, needs API key).--no-whisper-model— skip the auto-fetch of the Whisperbase.enmodel (~141MB). Default behavior downloads it so transcription works on first call.
A few of the tools need an extra click after the installer runs. The script prints these as it goes, but for reference:
| Tool | What you need to do |
|---|---|
| 21st.dev Magic MCP | Sign up free at 21st.dev/mcp (the MCP dashboard, not the homepage), grab your API key, follow the setup one-liner |
| Canva MCP | First time you call it, Claude opens a browser for Canva OAuth — allow, done |
| Figma MCP | First call opens browser for Figma OAuth — sign in with your Figma account, allow, done |
| Excalidraw MCP | First call opens browser for Excalidraw OAuth — sign in, allow, done |
| Gamma MCP | Opt-in. Default install skips Gamma (it fails to connect without a key). Re-run install with --with-gamma and grab a key from gamma.app/api |
| Morgen / Notion / Motion | Not in this repo — those live in cli-maxxing |
Everything else installs with zero accounts and zero keys.
| Tool | Kind | What it does |
|---|---|---|
| UI/UX Pro Max | Claude skill | Design reference library: 67 styles, 96 palettes, 57 font pairings, 13 stacks |
| Taste Skill pack | Claude skill | 8 variants that block generic LLM UI output |
| 21st.dev Magic MCP | MCP server | Natural-language component generator |
| Canva MCP | MCP server | Canva designs / brand kits / exports from the terminal |
| Higgsfield / Seedance 2.0 | Claude skills (15) | Style-specific video prompt engineers (cinematic, anime, product 360°, etc.) |
| Remotion best-practices | Claude skill | React-based programmatic video |
| YouTube Transcript MCP | MCP server | Transcripts from any public YouTube video |
| yt-dlp CLI | Homebrew | The binary behind yt-dlp MCP |
| yt-dlp MCP | MCP server | Pull audio/video from IG, TikTok, YT, and 1000+ sites |
| whisper-cpp | Homebrew | Local Whisper inference — no API key, runs offline |
| whisper-mcp | MCP server | Claude-side interface to whisper-cpp |
| ffmpeg | Homebrew | Video/audio glue (shared — prompted before uninstall) |
| Figma | MCP server (HTTP, OAuth) | Read Figma files, inspect frames, export design tokens, convert designs to code. Paste any figma.com URL and it resolves. |
| Excalidraw | MCP server (HTTP, OAuth) | Generate + edit Excalidraw diagrams conversationally. "Draw the architecture" → actual .excalidraw file. |
| Gamma | MCP server (opt-in via --with-gamma, needs API key) |
Generate presentations, docs, and landing pages from a prompt. Pairs well with UI/UX Pro Max context. Default install skips it (no key = silent connect failure). |
| Playwright MCP | MCP server | Microsoft's official browser automation — lets Claude log into and operate web apps with no API (Higgsfield, niche SaaS, anything you'd normally click through). |
Whisper base.en model |
~141MB binary at ~/.whisper/ggml-base.en.bin |
Auto-fetched from huggingface.co so transcription works on first call. Skip with --no-whisper-model. |
/copywriting skill |
Claude skill | Master-trained anti-AI-slop copy filter (Bernbach, Hegarty, Abbott, Trott, Wieden, Sugarman, Sackheim, Schwartz, Bencivenga, Gossage, Krone, McElligott). 20 files — 1 SKILL.md + 19 references covering voice library, frameworks, headlines, body copy, CTAs, proposals, psych triggers, compression, humanization, proofreading, quality gates. |
/watch skill |
Claude skill | Frame-aware video → structured study notes. Accepts a single video URL, a full YouTube channel URL, or a playlist URL. yt-dlp pulls the video, ffmpeg extracts scene-change frames, whisper.cpp (local, key-free — reuses the media module's ~/.whisper/ggml-base.en.bin) transcribes when captions are missing; Claude reads every frame as an image and writes ~/watch/library/<slug>/notes.md. Channel mode also writes index.md with cross-channel synthesis (recurring hooks, formats, cloneable moves). Groq / OpenAI Whisper still available as cloud fallbacks. 15 files: SKILL.md + 11 Python scripts + a SessionStart hook + slash-command entry point. |
All targets are idempotent — the install script checks and skips if already present. Figma, Excalidraw (and Canva above) install as remote HTTP MCPs — the script registers them; first tool call opens your browser for a one-time OAuth. Gamma is opt-in (--with-gamma) and requires an API key from gamma.app/api. Playwright installs as a local npx MCP — no credentials needed.
- UI/UX Pro Max — the encyclopedia. I don't call it directly much — it sits in context so Claude knows "these 13 stacks exist, these 96 palettes exist, these 57 font pairings work together." Lets me say "make it feel like Linear" and get something that actually looks like Linear.
- Taste skill pack — the anti-slop filter. Before Claude ships UI, the taste skill checks whether it's about to emit centered-hero-plus-three-card-grid. If yes, it rewrites. Different variants = different registers (see Taste skills below).
- 21st.dev Magic MCP — natural-language component generator. "Build me a pricing card with three tiers, popular tier highlighted" — Magic pulls from their library and returns actual JSX. I use this instead of letting Claude make up components.
- Canva MCP — the sleeper. Conversational design inside the terminal. I ask for banners, thumbnails, brand kit variants, exports — no tab-switching, no mouse. It's the fastest way to make three versions of anything and keep the one that lands.
- Higgsfield / Seedance 2.0 skills — 15 skills, each a specialized prompt engineer for a video style (cinematic, anime, product 360°, fight scenes, fashion lookbook, etc). I describe what I want, Claude hands back a paste-ready Seedance prompt with the 2-second hook, timeline, camera moves, and sound design already tuned.
- Remotion — the skill ships the Remotion team's best-practices doc so Claude writes idiomatic React video instead of reinventing timing every session.
- YouTube Transcript MCP — drop a YT link, get the transcript. No credential needed.
- yt-dlp + Whisper stack — this is the one I actually wanted most. Paste an IG Reel / TikTok / tweet link and Claude transcribes it locally using Whisper. No uploads, no paywall, no mangled timestamps. Works on anything
yt-dlpcan download (~1000 sites). /watch— the level-up on transcription. Instead of "give me the words," this is "give me the video." Paste a tutorial / lecture / creator video URL and Claude downloads it, picks scene-change frames with ffmpeg, lines them up against the transcript, then reads every frame as an image and writes section-by-section study notes — with the on-screen code blocks transcribed verbatim, the diagrams described, the hooks/structure tagged. The real cheat: paste a channel URL (youtube.com/@handle/videos) or a playlist with--limit Nand it loops the whole pipeline, then writes a top-levelindex.mdwith cross-channel synthesis — what hooks repeat, what thumbnail patterns work, what's actually cloneable. Runs key-free off the same local whisper.cpp the rest of this stack uses. Natural-language activation — you don't have to type the slash command; drop a URL with "watch this video", "study this lecture", "transcribe this reel", "summarize this with frames", "break down this video", "analyze this channel", "scrape this creator", "sweep this playlist", "frame by frame", "hook analysis", "script structure", "research this creator", etc., and the skill auto-triggers.- ffmpeg — the glue. Remotion needs it for rendering, yt-dlp uses it for postprocessing, Whisper uses it for audio conversion,
/watchuses it for scene detection + frame extraction.
/copywriting— the anti-AI-slop filter for words, same shape as taste-skill is for UI. Trained on the 12 names every serious copywriter cites: Bill Bernbach, John Hegarty, David Abbott, Dave Trott, Dan Wieden, Joe Sugarman, Maxwell Sackheim, Eugene Schwartz, Gary Bencivenga, Howard Gossage, Helmut Krone, Tom McElligott. Auto-activates when I ask for a headline, hero, body copy, CTA, manifesto, proposal copy, landing page, ad copy, brand voice, naming, or "rewrite this paragraph." Picks a voice + framework, drafts hot, then compresses 40-60% and runs a quality gate that bans the obvious AI tells (elevate,unlock,harness,leverage,journey,seamless,robust,delve, "In today's…", "Imagine a…", and the rest). Suspended in regular chat by/concisefromcli-maxxing— only takes over when the deliverable is going to a human audience.
The Taste Skill pack (Leonxlnx/taste-skill) ships 8 opinionated variants plus 3 global knobs. One skills add command expands it into 8 separate skill directories under ~/.claude/skills/ — each one is callable as a slash command by its installed name.
The variants:
| Slash command | When to use it |
|---|---|
/design-taste-frontend |
Default premium frontend rules + 3-knob tuner. Safe everywhere. |
/redesign-existing-projects |
Retrofit an existing project — audits current CSS first. |
/high-end-visual-design |
Premium agency-grade soft UI, spring animations. |
/full-output-enforcement |
Spec → finished build. Bans placeholder stubs. |
/minimalist-ui |
Editorial / Linear-Notion / typographic. |
/industrial-brutalist-ui |
Raw, mechanical, Swiss + terminal. Dashboards. |
/stitch-design-taste |
Google-Stitch-aligned DESIGN.md generator. |
/gpt-taste |
GPT-leaning variant of the taste filter. |
The 3 knobs (edit design-taste-frontend/SKILL.md to tune, 1–10 scale):
DESIGN_VARIANCE— 1-3 centered/clean · 8-10 asymmetric/modernMOTION_INTENSITY— 1-3 simple hover · 8-10 scroll-triggeredVISUAL_DENSITY— 1-3 spacious/luxury · 8-10 dense dashboards
Combos I actually use:
- Premium SaaS landing →
/design-taste-frontend+/high-end-visual-design - Cleaning up an old project →
/redesign-existing-projects - Editorial blog →
/minimalist-ui - Dashboard or data tool →
/industrial-brutalist-ui - Figma export → React →
/full-output-enforcement+/stitch-design-taste
Pull the latest without reinstalling from scratch:
bash <(curl -fsSL https://raw.githubusercontent.com/fidgetcoding/creativity-maxxing/main/update.sh)Every target is idempotent — anything already up-to-date is skipped.
bash <(curl -fsSL https://raw.githubusercontent.com/fidgetcoding/creativity-maxxing/main/uninstall.sh)Or, from a clone:
bash uninstall.shRemoves, in reverse order: UI/UX Pro Max, all 8 taste variants, 21st.dev Magic MCP, Canva MCP, Figma MCP, Excalidraw MCP, Gamma MCP, Playwright MCP, Higgsfield/Seedance skills, Remotion skills, YouTube Transcript MCP, yt-dlp MCP, yt-dlp CLI, whisper-cpp, whisper-mcp, /copywriting skill, /watch skill, watch config + library cache at ~/.config/watch/ + ~/watch/library/, ffmpeg. The shared whisper model at ~/.whisper/ggml-base.en.bin is only touched if you accept the whisper-cpp brew uninstall prompt. Prompts before removing ffmpeg because ffmpeg is usually shared with non-creative tooling — answer N to keep it.
This is one of three repos in the stack:
| Repo | What it does |
|---|---|
cli-maxxing |
Foundation — Claude Code, shell aliases, dev tools, productivity MCPs (Morgen, Motion, n8n, Notion, Playwright, SwiftKit). |
creativity-maxxing |
This repo — design skills, video prompt engines, transcription lab, Canva in terminal. |
task-maxxing |
Two-way task sync — Obsidian ↔ Morgen (Notion dropped 2026-05-04) (requires 2ndBrain-mogging). |
Install cli-maxxing first (it drops claude onto your PATH). After that, creativity-maxxing and task-maxxing can be installed in either order.
See SECURITY.md — credential surface, supply-chain notes, and how to report installer vulnerabilities.
TL;DR: no secret ever goes into this repo, no install step downloads an unsigned binary over HTTP.
MIT — see LICENSE.
This repo ships with a .gitleaks.toml config and a one-liner installer for a local
pre-commit hook that scans staged content for secrets (GitHub tokens, API keys, JWTs,
etc.) before every commit.
bash scripts/install-pre-commit-hook.shThe hook runs gitleaks protect --staged and blocks commits that contain secrets.
For emergencies you can bypass with git commit --no-verify — but DO NOT bypass for
real secrets. Use env vars or a secret manager instead.
If gitleaks isn't installed yet:
- macOS:
brew install gitleaks - Linux: https://github.com/gitleaks/gitleaks/releases