Studio Setup

About 15 minutes, once. At the end the creator has:

• the tools the other skills call, installed and checked,
• video-agent/voice.md: who they are, how they talk, what they believe, what they can prove. Every script is written from it,
• video-agent/styles/default/style.json: their look on top of the split-screen paper layout (their color, fonts, captions, paper, pace, sound, Instagram CTA), with the real logos of the tools they talk about. Every edit uses it,
• the render engine ready (Chrome, whisper.cpp, the music beds), so the first edit starts right away,
• permission rules, so Claude can run ffmpeg, yt-dlp and the engine without asking each time,
• if they want: their accounts connected through Treg (audit and publishing), and the keys for translation and images.

How to run it

• Talk in the creator's language. Simple words, no jargon.
• One question per message. Offer choices when you can, recommended first.
• Never invent anything about them: a fact you don't have becomes a question or a {{placeholder}}.

1. Tools

Run the checks and show one line per tool (✓ version, or ✗ missing):

node -v
ffmpeg -version
yt-dlp --version
deno --version
git --version

Give the commands for what is missing and let them run them (or run them yourself if they say so). Scripting-only users (studio-script, its viral mode included) only need yt-dlp and deno. Codex is only for /codex-images and needs a ChatGPT plan; skip it if they don't want images.

• Windows: after winget, close and reopen the terminal (and Claude Code), or the new commands are not found.
• YouTube changes often. When yt-dlp fails on a link that plays in the browser, update it first (brew upgrade yt-dlp, winget upgrade yt-dlp.yt-dlp, pipx upgrade yt-dlp).

2. The folder

Create it at the root of the current project:

video-agent/
├── voice.md      who they are and how they talk (step 4)
├── swipe.md      hooks that worked in their niche (`/studio-script viral`)
├── audit.md      what works on their own Instagram (insta-audit)
├── audits/       full Instagram reports, one folder per account and run (insta-audit @handle)
├── log.md        one line per script written, so nothing repeats
├── scripts/      one file per week (studio-script)
├── takes/        raw recordings (phone, camera, webcam, screen)
├── style.json    {"default": "default"}: the style every edit uses (step 5)
├── styles/       default/ (their look, step 5), and styles copied from other creators (steal)
├── edits/        one folder per edited video (short-editing): source/spec.js, output/<name>.mp4 + post.md, cover.jpg, published.md
├── repurpose/    one plan per long video (clips)
├── images/       generated images not tied to a video (codex-images)
├── .env          their API keys, never shared (step 7)
└── .tmp/         scratch files, deleted after each run

log.md starts with # Script log, swipe.md with # Swipe file.

If the project is in git, add these lines to its .gitignore (videos are heavy; the text files stay tracked):

video-agent/takes/
video-agent/edits/**/*.mp4
video-agent/edits/**/assets/
video-agent/edits/**/stills/
.claude/skills/short-editing/.runtime/
.claude/skills/short-editing/engine/node_modules/
video-agent/.tmp/
video-agent/.env

video-agent/.env holds keys: it must be ignored even if they skip the rest.

3. The engine

The render engine ships inside the short-editing skill (<skill> = that skill's folder, e.g. .claude/skills/short-editing). Its first setup installs the engine packages (15 MB), a Chromium when no Chrome is installed (100 MB), whisper.cpp and its speech model (0.5 GB) and the 8 music beds: 3 to 6 minutes, once. Ask: now (recommended if they will edit soon, and it lets step 5 show their style on their own footage), or later (short-editing runs it the first time). Only for short-editing, clips and short-translate; scripting-only users skip it.

node <skill>/engine/setup.mjs --check || node <skill>/engine/setup.mjs

Say what is downloading and how big before starting. ffmpeg missing: it prints the install command.

4. voice.md

First the setting at the top of the file: the language they record in (their brand color is asked in step 5). Then ask how to build the rest:

1. From their videos (recommended when they have posted before): 3 to 5 links of their best talking videos (YouTube, Shorts, TikTok, Instagram reels). Pull the transcripts (<lang> = their language code: en, fr, es):

   yt-dlp --skip-download --write-auto-subs --write-subs --sub-langs "<lang>.*" --sub-format vtt -o "video-agent/.tmp/%(id)s.%(ext)s" "<url>"
   

   Copy
   Read them and fill the template from what they actually say: their real phrases, their pace (words ÷ minutes of speech), the words they repeat, the way they open and close. Then ask only what the transcripts can't tell (positions, proof, off limits, the ask). Delete video-agent/.tmp/ after. YouTube links almost always give a transcript. TikTok sometimes does, Instagram almost never: for those, ask for a YouTube link or switch to the interview.
2. From an interview: ask the template's questions one at a time, in plain words ("What do you believe that most people in your niche get wrong?").

Either way, ask them to paste 3 short texts they wrote themselves, before any AI tool (a caption, a message, an email). They go in "Real samples" and teach the scripts their rhythm.

Write video-agent/voice.md from references/voice-template.md. Then read it back to them in 5 lines and ask what is wrong. A voice file is never finished: say they can run /studio-setup voice any time to update it.

5. Style

Five minutes. Every edit has the same finished layout, the split-screen paper look: boards on a textured paper above the camera card, hard cuts between split, paper only and full camera, captions on the seam, a comment-sheet CTA (style.md). This step makes it theirs: video-agent/styles/default/style.json. The format and the five tested looks: looks.md; every key: style-schema.md. Read both first.

Say it in one line before starting: "The default already looks finished; I'll ask 7 quick questions to make it yours, then show it on your own video." Then one question per message, choices with the recommended one first, and their answer can always be "default" or "skip". Use what voice.md already says (niche, tone, the ask) to pick the recommendation and to skip a question it already answers.

1. Look: show the five looks of <skill>/assets/looks.json in one line each (label + says): default (minimal: the system font, the camera full width, mono caps captions on a black band on the seam, recommended), classique (sticker captions), papier (framed camera photo, the said word on a color block), pop (podcast, opinion), poster (fast, hard cuts). Their pick's style patch is the start of the file.
2. Accent color: "Your brand color? A hex code, or a color in words." They name a color: turn it into a hex → colors.accent. None: keep the look's. It marks the accent words, the caption block, numbers, so it must be strong (not a pastel).
3. Fonts: "Keep the system font (the one of the Mac, San Francisco; Segoe UI on Windows, recommended), or Bricolage Grotesque (modern, tight), Poppins (round, friendly), Archivo (heavy, condensed), Inter (neutral)?" The system font is system-ui, -apple-system, Segoe UI, Helvetica Neue, Arial, sans-serif. → font.family, display.font / display.text. Their own font file: copy it to styles/default/assets/fonts/ and set font.src.
4. Captions: "Band: mono capitals, white on a black rounded band on the seam, 3 words at a time (band + JetBrains Mono, recommended), minimal white words on the video, each one appearing when said (rise + block: false + straddle + straddleDy: 90), the said word on a color block (rise), hard shadow (hard) or sticker?" Then "Normal size, or bigger?" (bigger = captions.size 96). UPPERCASE = captions.upper: true.
5. Rhythm: "Fast (1.15x, the default), Calm (1.05x) or Natural (1x)?" → pace.speed.
6. Instagram CTA: "Your Instagram handle, and the word you ask people to comment (GUIDE, LINK, the product name)?" → creator.handle, creator.name, and a note "Comment word: <WORD>" in notes. Also creator.lang (their recording language) and creator.niche from voice.md.
7. Tools you talk about: "Which apps or tools do you name most in your videos?" Up to 8. For each, fetch the real logo (looks.md, "Logos"), save it as video-agent/styles/default/assets/logos/<tool>.png. A logo you can't find for sure: ask them for the file, never draw a look-alike.

Write video-agent/styles/default/style.json ("version": 1, "name", the look's patch, then their answers; only keys that differ from the engine defaults) and video-agent/style.json = {"default": "default"}. Check it parses: node -e 'JSON.parse(require("fs").readFileSync("video-agent/styles/default/style.json","utf8"))'.

Preview, when the engine is set up (step 3) and they have a take of 10 s or more in video-agent/takes/:

node <skill>/engine/short.mjs init video-agent/edits/short-style-test --source "<take>"

Then in its source/spec.js add one board around second 1 to 4 of the take (a text with their accent word, their first tool's logo; scene-library.md), and:

node <skill>/engine/short.mjs words video-agent/edits/short-style-test
node <skill>/engine/short.mjs prepare video-agent/edits/short-style-test
node <skill>/engine/short.mjs stills video-agent/edits/short-style-test 0.5,2,5

Read the JPGs in stills/ yourself first: the accent readable on the paper, white caption text readable on its block, the logo crisp, nothing cut at the edges. Fix what fails, then show them and ask: "Keep it, or change one thing?" A change = edit style.json, init again (it snapshots the style), new stills. Trash short-style-test after. No take yet: describe the look in 3 lines, the first edit shows it.

They can run /studio-setup style any time to change it. A creator's style copied with /steal sits next to it (video-agent/styles/<name>/) and is used with /short-editing <take> style <name>.

6. Permissions

The other skills run the same few commands again and again. Pre-approve them for this project. Show the rules, ask, then merge them into .claude/settings.json. Create the file if it is missing. Keep every key and rule already there, and never remove anything.

{
  "permissions": {
    "allow": [
      "Bash(yt-dlp *)",
      "Bash(ffmpeg *)",
      "Bash(ffprobe *)",
      "Bash(node .claude/skills/short-editing/engine/*)"
    ]
  }
}

• Only these. Never add Bash(*), Bash(rm *), Bash(curl *) or a rule for a whole interpreter (Bash(node *), Bash(python *)).
• Rules in the shared settings.json apply once the folder is trusted. A creator who keeps this project in git and doesn't want to share the rules can use .claude/settings.local.json instead (same format, personal).
• Each skill also lists its own commands in its allowed-tools, which covers the turn the skill runs in. These rules cover every other turn.

7. Accounts and keys (optional)

Three things, each only if they want the skill that needs it. Ask once, in one message:

ElevenLabs: they create a key at elevenlabs.io/app/settings/api-keys and paste it themselves into video-agent/.env as ELEVENLABS_API_KEY=... (create the empty file and open it for them). Never ask for a key in the chat. Check it is there without printing it: grep -c ELEVENLABS_API_KEY video-agent/.env.

Codex: codex --version, then they run codex login (browser, "Sign in with ChatGPT").

Treg

yt-dlp reads YouTube for free. TikTok and Instagram hide their numbers from scripts, a creator's own stats sit behind their login, and posting needs their accounts. Treg fills that gap: their own accounts connected once (stats and publishing, free), other creators' public numbers for about $0.001 a call. New verified accounts get $1.00 of credit once, enough for hundreds of calls.

Two ways to plug it in:

• CLI, macOS and Linux. It needs uv (recommended, it fetches the right Python), or pipx or pip with Python 3.12 or 3.13. There is no Windows installer.

  curl -fsSL https://treg.to/install.sh | sh
  treg login
  

  Copy
  treg login opens the browser (GitHub, Google or an email code).
• MCP, any OS, nothing to install: claude mcp add --transport http treg https://treg.to/mcp/, then /mcp to sign in. Each treg catalog search, catalog get, call and balance below becomes the MCP tool catalog_search, catalog_get, call or balance.

Their own accounts (connecting needs the CLI). --capability post asks for the rights to read their stats and to post; posting still waits for their yes on each video:

treg connections connect --provider instagram --capability post
treg connections connect --provider tiktok --capability post
treg connections connect --provider youtube --capability post
treg connections ls

Each connect opens the platform's own login; they approve it there. Connect only the platforms they post on. Instagram needs a professional account (Creator or Business, a free switch in the app: Profile → menu → Account type and tools). Treg prints tips on stderr: add 2>/dev/null when you read its JSON.

Instagram connected: offer /insta-audit right after, it is the best first look at what works for them.

How every Treg call works, the rule for all the skills:

1. treg catalog search "<what you want to do>": describe the job, not a vendor ("list a TikTok account's videos with view counts").
2. treg catalog get <id>: the parameters and the price. Say the price before calling. Their own connected accounts are marked free.
3. treg call <id> --query key=value. A failed call is not charged. treg balance shows what is left; a 402 error means the balance is empty.

Permissions: add only the free, read-only commands, so every paid treg call still asks first:

"Bash(treg catalog *)",
"Bash(treg balance)",
"Bash(treg connections ls)"

8. Wrap up

Show what is ready (✓ or skipped, one line each) and the path forward. They can also just type /editing <what they want>: it picks the right skill.

The shortest path to a first video: record one take (phone, camera or webcam) into video-agent/takes/<slug>.mp4, then /short-editing. Then, when they want more:

1. /studio-script: this week's scripts, hooks first, humanized (/studio-script viral first for what works in their niche).
2. /insta-audit: what works on their own Instagram (fills audit.md).
3. /outliers: the reels that blew up for their competitors, turned into 5 scripts to record.
4. /short-editing: a short (vertical) or a full YouTube video (horizontal), in their style.
5. /broll: their own B-roll clips as a library, placed on their edits.
6. /steal: a creator's reel they love into a second style.
7. /clips: a long video into finished shorts.
8. /short-translate: a take in another language, in their voice.
9. /codex-images: thumbnails, covers, end cards.
10. /vox: a photo into a paper-collage animation.
11. /publish: YouTube, Instagram and TikTok.

Later

/studio-setup <part> jumps to one part: tools, engine, voice (rebuild or update voice.md), style (their look, sound, CTA, tool logos, logo, end card), permissions, treg, keys.