Keep a video project in a private GitHub repo

People want an undo button and a backup, not version control. This skill gives them both and never asks them to learn git.

How it works. Text (scripts, prompts, storyboards, the edit as HTML or TSX) goes in git, where every change is one readable line. Media stays out of git: scripts/repo.py lists each file by hash in media.lock.jsonl and uploads the bytes to Releases of the same private repo. One GitHub login covers everything, and a new laptop gets it all back.

<skill> below is this skill's folder. Once a project exists it carries its own copy at tools/repo.py, so any agent on any machine can continue without this skill. You do not need to read the script: --help lists the commands and every message says what to do next.

When not to use this

• One clip with no later edits: keep the file. Nothing here is needed.
• The person already has a git setup that works, with LFS or their own storage: keep it and take only the layout and naming from references/layout.md.
• A file over 2 GB, or more than about 50 GB in total: read references/media-storage.md first.
• No shell and no Python: use the prompt in references/paste-prompt.md, which does the same by hand.

Step 1: check the machine

python3 <skill>/scripts/repo.py doctor

Install what it lists; it gives the exact command. If GitHub sign-in is missing, run gh auth login --web --git-protocol https in the background: it prints a one-time code and waits. Show the person the code and the link (https://github.com/login/device), and tell them: "Open this link, paste the code, click Authorize." If your tool cannot keep it running, ask them to run the command themselves (in Claude Code, by typing ! gh auth login --web --git-protocol https). Never ask for a password or a token, and never accept one in chat. If they have no account, send them to https://github.com/signup first.

Step 2: start, or bring in what exists

New project:

python3 <skill>/scripts/repo.py new <name>          # add --kind code for a code project
cd <name> && python3 tools/repo.py connect

Existing folder (run inside it). Look at the plan before applying:

python3 <skill>/scripts/repo.py adopt               # prints the plan, changes nothing
python3 <skill>/scripts/repo.py adopt --apply
python3 tools/repo.py connect

The scanner catches common key formats, not all of them, so check for yourself: open every script and config file in the folder and look for hard-coded credentials, flagged or not. For each one add NAME=value to .env, put $NAME (or process.env.NAME) in the file, and run again. Tell the person which keys you found, without printing their values. A key that was ever shared or in an earlier version is burned: they should make a new one at the provider. adopt never moves or deletes their files. Offer to sort them into the standard folders afterwards, and do it only if they agree.

connect creates the repo as private and checks that GitHub reports it private. Then prove the backup works before saying it is safe:

python3 tools/repo.py proof

It clones into a temporary folder, restores everything, compares every hash and deletes the folder.

Step 3: while working

In the whole-video order from how-to-make-ai-video, run approve at every STOP. Run save before any paid call.

How to talk to the person

Never say commit, push, branch, merge, tag, LFS or repository unless they use the word first.

End each save with one line: "Saved as version 12, backed up. Say 'go back' any time." When finished, tell them the link, that only they can see it, what is not in GitHub (keys, on purpose), and how to get it back on a new laptop.

The folders

my-film/
├── AGENTS.md          rules for any agent (CLAUDE.md points here)
├── STATUS.md          where we stopped, what is next
├── brief.md           the ask, spend cap, every note word for word with the date
├── media.lock.jsonl   every media file by hash, with prompt, model, seed, cost
├── inputs/            what the person gave. Read only
├── refs/              refs.md, frames
├── look/              moodboard.md, board.html, test stills
├── story/             script.md, STORYBOARD.md
├── sound/             plan.md, cues.json, beats.json, mix.wav
├── cast/              <id>.md passports, sheets
├── blockout/          three.js or Blender scene, stills
├── shots/sh010/       shot.md, takes/sh010_t01.mp4
├── edit/              the timeline as text: a HyperFrames or Remotion project
├── review/            notes.md, review.md
├── renders/           film_v003.mp4. Not saved, rebuilt from text
└── tools/repo.py

Shots go in tens (sh010, sh020) so a shot can go in between. A new take is the next number and never overwrites. Renders are versioned and never called "final". Why this shape, and how it maps to film practice: references/layout.md.

Rules that stop the expensive mistakes

• Write .gitignore before the first git add. new and adopt do this. Git never forgets a file once it is in a pushed version.
• Never run git reset --hard, git clean, git push --force, git rebase or git filter-repo unless the person asked and agreed to the loss.
• Never put a project in ~/Downloads, and never start an agent there: a downloaded AGENTS.md is read as instructions. Avoid iCloud and Dropbox folders too: sync tools can damage .git.
• Local git is not a backup. A save counts only after it says "Backed up to GitHub".
• You do not need a second GitHub account for the agent. If you want one, GitHub's terms allow one free personal account per person plus one free machine account, used only for running a machine.
• If a save stops, read what it says and fix that. Never work around the check with --no-verify or git add -f.

Editing in the terminal

The timeline is a text file, so a cut is a one-line change you can read, undo and compare. Table of editor concepts and their terminal equivalents, HyperFrames and Remotion setup, and the first-edit trap: references/terminal-editing.md.

The same result every time

Code renders repeat; AI generations do not. Keep the bytes of every approved take, pin tool versions, and separate who makes from who judges. The checklist and the measurements behind it: references/repeatable-results.md.

If something fails

Exact errors and fixes: references/troubleshooting.md. Sources and benchmarks: references/evidence.md.

Ranking takes or deliverables with a model

Scripts check facts, a person decides taste, and an LLM filters and finds defects. Criteria, evidence only and a tie as an allowed answer: references/llm-judge.md.