Skip to content

TsukiNi22/ai-utils / skills/studio/setup-studio

setup-studio

Unverified

Set up the creative studio of the user's computer - scans the tools there for audio (music, sound effects, voice), image, video, 3D (modeling, CAD) and simulation / VFX / rendering (programs, flatpaks, Python packages, MCP servers, claude.ai connectors, GPU), asks what to cover and whether the work is commercial, searches online for free tools without subscription for what is missing (license and commercial use checked, MCP servers too) and installs the accepted ones, then generates a /studio router and one studio-audio / image / video / 3d / vfx skill saying which installed tool to use for each task, where, how (MCP, command line, GUI) and why, with commands and license notes - in ~/.claude/skills or as a claude.ai personal skill; run again it refreshes them. Use whenever the user wants to set up, scan or refresh their creative / media / audio / video / 3D / VFX tools or MCP servers, or summons /setup-studio.

Setup studio: the creative tools of this computer, as skills

SKILL_DIR = the directory of this file. Like setup-writing-style, this skill builds other skills: it finds what the computer has for creative work, fills the gaps with free tools the user accepts, then writes a /studio router and one skill per domain that says, for each task, which tool, where, how and why.

DomainGenerated skillRoles (reference/catalog.json)
Audio: music, sound effects, voicestudio-audioedit, compose, synth, sfx, process, stems, voice, transcribe, generate, assets
Image: 2D, retouching, painting, vector, generationstudio-imageedit, paint, vector, photo, batch, optimize, generate, upscale, cutout, vectorize, assets
Video: editing, effects, motion, generationstudio-videoedit, cli, compositing, motion, record, subtitles, transcode, generate, download
3D: modeling, CAD, texturingstudio-3dmodel, cad, mesh, texture, lowpoly, generate, realtime, assets
Simulation, VFX and renderingstudio-vfxsim, render, compositing, cfd, physics, realtime

assets/ holds the studio's look, copied into what is generated: the logo (assets/logo.svg flat for any size, assets/logo.png, assets/logo-3d.png rendered in Blender: a glass prism splitting a ray into the five domains under a clapper) in the router, and one icon per domain (assets/icons/<domain>.svg / .png) as assets/icon.svg of each domain skill (all of them in the combined claude.ai skill).

The catalog lists ~70 free tools with their roles, detection, license, commercial use (yes / conditions / no), install commands, MCP servers known, GPU needs and ready commands. It is the starting point, not the limit.

0. Where it runs

  • Claude Code on the user's computer (a shell there): everything below. Work files in a temp folder (WORK=$(mktemp -d), or the job's tmp folder).
  • claude.ai / no shell on the user's computer: the sandbox is not their computer, never scan it. Ask what they have per domain (step 2, options taken from the catalog) and use the connectors of the session; then inventory.py --declare <ids> --connectors <names> makes the inventory (its hardware lines are the sandbox's: ask the GPU, put it in the notes) and step 6 saves one combined skill.
  • $HOME/.claude/skills/studio/reference/state.json exists: refresh (step 7).

1. Scan

python3 SKILL_DIR/scripts/inventory.py --json "$WORK/inventory.json" --connectors <Name,Name>
claude mcp list                     # also the MCP servers of plugins, and whether each one answers
  • --connectors: the claude.ai connectors of this session, read from the tool list (mcp__claude_ai_<Name>__*).
  • The summary gives, per domain, what was found and the roles nothing free covers; the JSON has the paths, versions, flatpaks, MCP servers (never their environment or tokens), GPU / VRAM, RAM, free disk, installers.
  • An MCP server or a program the catalog does not know but that looks creative: note it for step 2 (its domain).

2. Ask (interactive-questions; clarify for real doubts, ideation when the user wants proposals)

Never ask what the scan answered. Round 1, one popup:

  1. Domains to cover (multi, all five recommended).
  2. Commercial use: Oui, travail commercial (the tools marked no are left out) · Non, usage perso · Selon le projet (asked each time, commercial-safe tools first). Saved as commercial: yes / no / ask.
  3. Installs allowed (multi): system packages (root, the user's way: SUDO_ASKPASS=... sudo -A dnf ...), Flatpak (--user, no root), Python tools (uv tool install, isolated), MCP servers, AI models (big downloads), none.
  4. Local AI generation: show what the GPU allows (VRAM: 4 GB = SD 1.5 / SDXL with offloading, Whisper small, no local video; 8 GB = FLUX, Stable Audio Open; 12 GB+ = video models) - yes / no / only the light ones.

Round 2, only where it matters: several installed tools for the same role and the choice changes the work (two video editors, two DAWs): which one first (prefer, ex: video.edit = kdenlive). An unknown MCP server: its domain.

3. Fill the gaps (online search)

For each domain chosen and each missing role (and roles covered only by a connector spending credits):

  • Start from the catalog candidates (free, commercial-compatible when needed, fit for the GPU and the free disk).
  • Search online (WebSearch) for better or newer free options: best free open source <role> linux <year>, <tool> license commercial use, <tool> MCP server. Keep a tool only when checked at its source (official site, LICENSE of the repository): free, no subscription (freemium / credits said as such, never proposed as the main choice), installable on this system (dnf / Flatpak / pip / AppImage), released in the last two years. An MCP server: its repository read (license, activity, what it runs, local only, no paid API key unless the user wants it).
  • Present them per domain in a QCM (multi): label = name, description = what for, license, commercial yes / no / conditions, how it installs and its size, GPU; the recommended first; Rien pour ce domaine. What is refused goes to declined (not proposed again at the next refresh unless the user asks).

4. Install (only what was accepted)

  • Packaged in the distribution: dnf (root, the user's way); else flatpak install --user -y flathub <id>; Python tools uv tool install <pkg>; models in a folder the user names, the size said first. Check each name before (dnf info, flatpak remote-info flathub <id>), never curl | sh from an unknown site.
  • MCP server: claude mcp add -s user <name> -- <command> (or --transport http <name> <url>), then claude mcp list until it answers; its own steps too (ex: blender-mcp needs its addon.py installed in Blender and Connect clicked). It works in the next session.
  • After each install: its --version or a smoke run; a failure is reported as it is, never hidden. Then scan again (inventory.py --json).

5. Choices

Write $WORK/choices.json from the answers:

{"commercial": "yes", "domains": ["audio", "image", "video", "3d", "vfx"],
 "prefer": {"video.edit": "kdenlive"}, "skip": [], "declined": ["houdini-apprentice"],
 "mcp": {"my-daw": "ardour", "other-server": "domain:audio"},
 "howto": {"kdenlive": "MCP kdenlive: open_project, add_clip, render ... (read with ToolSearch)"},
 "extra": [{"id": "x", "name": "X", "kind": "cli", "domains": {"audio": ["sfx"]}, "where": ["bin:/usr/bin/x"],
            "license": "MIT", "commercial": "yes", "why": "...", "recipes": [["what", "command"]]}],
 "notes": {"audio": "- a first note of the user"}}

howto = what the scan cannot know (the real tools of an MCP server: load them with ToolSearch and name the main ones); extra = a tool found online and installed that is not in the catalog (its license checked).

6. Generate and save

python3 SKILL_DIR/scripts/generate.py --inventory "$WORK/inventory.json" --choices "$WORK/choices.json"

Writes $HOME/.claude/skills/studio (router /studio, its state.json for the refresh) and studio-<domain> per domain, from templates/router.md and templates/domain.md. It never overwrites a skill it did not generate (--force only after asking), keeps the notes between the studio:keep markers, removes the generated domains the user dropped; --dry-run shows what it would write. Then read every generated file and fix what the scan could not know (a wrong pick, a tool used another way here): change choices.json and generate again rather than editing by hand.

Save the way setup-writing-style does, by what this session has:

  1. Claude Code: the files above, nothing else to do (new skills are listed from the next session at the latest).
  2. claude.ai with a save_skill tool: generate.py --single --out "$WORK/single" (one skill studio with every domain), then save_skill with name: "studio", its description and the body (everything below the frontmatter); overwrite: true only when a studio skill is already listed (its exact name). One skill: the quota is spared.
  3. claude.ai / Cowork without it: generate.py --single --zip "$WORK/zip" and give studio.zip to upload (Customize > Skills > Upload); from Claude Code too when the user wants the same studio on claude.ai (the programs then run on their computer, the connectors in claude.ai). The state.json is never put in a zip.

7. Refresh (run again)

python3 SKILL_DIR/scripts/inventory.py --json "$WORK/inventory.json" --connectors <Name,Name> \
    --compare "$HOME/.claude/skills/studio/reference/state.json"

Start from the choices of state.json: ask only about what changed (tools added or removed, new missing roles, a domain to add); declined stays declined. Then step 6 again (the notes are kept).

8. Report

In the user's language, short: per domain the tools used (MCP / command line / GUI), what was installed now, what is still missing or declined, the licenses that limit commercial use, where the skills are (or the save still pending), and how to use them: /studio <request> or simply the request (the domain skills load by themselves).

Rules

  • Nothing installed, no MCP server added, no model downloaded without a yes; nothing paid or with a subscription proposed as a solution (only said when nothing free exists).
  • A license is never guessed: unknown = check, said to the user.
  • No secret in the files: the MCP environment, headers and URL queries are never read into them.
  • The generated skills describe this computer only: never committed to a repository.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 464kUpdated Claude CodeCodexVerifiedInstall
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 464kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 318kUpdated Claude CodeCodexUnverifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 318kUpdated Claude CodeCodexUnverifiedInstall