Zkitszo - The Real
October 18, 2026

Google Drive on Steroids- GitHub Without the Code

Months of chats, custom instructions, skills, a little website spun up in a vibe-coding app... all of it living inside somebody else's product. Then the work account gets cancelled, or a shinier assistant shows up- OpenAI's dots landed at the end of September 2026, another one is surely on the way- and every bit of that context is stuck where it was made.

GitHub is the way out, even for folks who never plan to write a line of code. Picture "Google Drive on steroids", built on Git. A lot of the reading and writing ends up done by the AI.

Where the Work Lives Your Computer AI agent, skills, notes, site files (.gitignore keeps keys out) GitHub Repo set to private every commit is a version to roll back to Live Website Cloudflare or Vercel deploys on push Any AI Assistant reads the repo as its memory New Computer git clone, carry on commit + push daily, on a timer The repo stays put if any one app gets cancelled.

Four Words, Maybe Five

The jargon scares people off. GitHub's own glossary starts right where a Drive user would:

A repository is the most basic element of GitHub. They're easiest to imagine as a project's folder.

A repository- "repo" for short- is the Drive folder.

A commit is one saved change. Same idea as an entry in Google Docs' version history.... every commit is a point we can roll back to.

A push sends those commits up to GitHub- the upload. A pull request is a proposed change somebody accepts or rejects- suggesting mode in Docs by another name. A fork is a personal copy of somebody else's repo, sitting on our own account.

GitHub's free Introduction to GitHub course promises the basics "in less than an hour". The Markdown these files are written in gets its own walk-through in Get Up Get Down Get Hip With Git Markdown.

Free, with Fine Print

GitHub Free is "$0 USD per month forever" with "Unlimited public/private repositories". Private repos get what GitHub calls a "limited feature set".

For single files, Git warns at 50 MiB, GitHub blocks anything over 100 MiB, and the upload button in the browser tops out at 25 MiB- videos and model files go somewhere else.

GitHub's own docs disagree on whole-repo size: "ideally less than 1 GB, and less than 5 GB is strongly recommended" on the large-files page, 10 GB on disk on the newer repository-limits page. Notes, prompts and skills sit far below either.

Back Up the Whole AI Setup, Daily

Claude Code, Codex, Cursor, the desktop chat apps... the local ones keep skills, instruction files and notes on one computer. One dead SSD takes all of it! A private repo with a fresh commit every day means a new laptop just clones it and carries on where the old one stopped.

Keep the secrets out before the first commit. GitHub's free secret scanning runs on public repos. The push protection switched on for every account by default "Stops you from pushing secrets to public repositories". A private backup repo gets neither by default.. an API key that slips in just sits there in the history. A .gitignore file- a plain list of names Git should pretend not to see- keeps keys out, along with any folder holding client work under NDA. If a key gets committed anyway, GitHub's advice is to "revoke and/or rotate that secret" first. Deleting the file afterwards doesn't pull it back out of the old commits.

The gh tool is GitHub's command-line helper (brew install gh on a Mac). GitHub Desktop does the same jobs with buttons, free, on macOS and Windows. In a terminal, from the folder the agent works in:

gh auth login
cd ~/ai-setup
git init -b main
printf '.env\n*.env\n*.key\n*.pem\nsecrets/\nnda/\n' > .gitignore
git status
git add -A
git commit -m "first backup"
gh repo create ai-setup --private --source=. --remote=origin --push

git status lists what's about to be saved- no key files should show up in it. Then the backup itself, saved as ai-backup.sh in the same folder. It only commits when something actually changed:

#!/bin/sh
cd "$(dirname "$0")" || exit 1
git add -A
git diff --cached --quiet && exit 0
git commit -q -m "backup $(date +%F)"
git push -q -u origin main

Cron is the timer built into macOS and Linux (Windows has Task Scheduler). crontab -e opens its schedule. This line runs the backup at 3 pm every day:

0 15 * * * /bin/sh "$HOME/ai-setup/ai-backup.sh" >> "$HOME/ai-backup.log" 2>&1

Run it once by hand first and look for the dated commit on GitHub. The log catches anything the timer trips over.

A Briefcase Any Assistant Can Open

The same trick works as memory for a personal assistant. Have it write a dated Markdown log into a repo every day- what it worked on, what it learned, what's on deck tomorrow. Point the next assistant at that repo with something like "here's my history from the last assistant, pick up where it left off".

Whatever an assistant remembers lives wherever that assistant keeps it. The repo is the copy that comes along to the next one. One write-up, Your AI App Should Not Own Your Work, calls it the binder method: an AI app checks the binder out, works in it, checks it back in. Its companion prompt builds the folders- a today file, projects, agent instructions, skills.

Keep it lean... piling thousands of files in front of a model can make answers worse ("context rot"). A short index file pointing at the rest beats a dump.

Claude connects to GitHub repos directly. ChatGPT's GitHub app "only lets you read from your repositories"- writing back needs Codex, or an agent with Git on a real computer.

Websites Off the Monthly Builder

Whole websites run the same way. An agent like Claude Code or Codex writes the pages, GitHub stores them, a host builds and publishes them on every push. Cloudflare's GitHub integration "will automatically deploy your code every time you push a change". Astro, a framework "optimized for building fast, content-driven websites", suits a blog.

Cloudflare now tells people to "Start new projects with Workers" instead of its older Pages product. Pages still works: 500 builds a month on Free, 20,000 files per site, 25 MiB per file.

Vercel's Hobby plan is free as well. Its fair-use page restricts Hobby to "non-commercial personal use only". Ads such as AdSense count as commercial, as does a site that's mostly affiliate links.

A warning goes around that cancelling a builder like Lovable takes the website down with it. Lovable's own docs say otherwise: cancelling drops the workspace to Free. "Your published apps stay live on Free." Git sync to GitHub works on every plan. Custom domains are paid-only. A Free project with no visits and no owner login for 12 months can be unpublished, code kept. The warning is wrong. Syncing to GitHub on day one is still worth it.

Borrowed Repos

Running a public repo raw rarely fits- forking it, or handing the link to an agent with "build our own version on these principles", gets something that does. A skill is just a folder with a SKILL.md file in it, read the same way by Claude Code, Codex, Gemini CLI and others (more on that in Usage Limit Reached? Not So Fast..).

  • anthropics/skills- Anthropic's own collection, with skill-creator, frontend-design and mcp-builder. No single licence: many are Apache 2.0, the docx, pdf, pptx and xlsx skills are source-available only.
  • HyperFrames- HeyGen's framework that turns HTML, CSS and animations into MP4 video, Apache 2.0 with "no per-render fees". The Studio desktop app is out for macOS and Linux; its page names no price.
  • Google Workspace CLI- one command-line tool for Drive, Gmail, Sheets, Docs and the rest. Its README says it is "not an officially supported Google product".
  • Suganthan's GSC MCP- an MCP server (a plug that lets an AI app call an outside tool) for Google Search Console data. "Read only" holds only if chosen: the setup wizard asks for view access, but a manual config without GSC_SCOPES=readonly gets full access, URL-submission tools included.
  • REA- "Reverse engineer anything with agents", MIT licensed, meant for software we own or are authorized to examine. It showed 318,399 stars six months after it was created.. real fans, or bought stars?

Read Before You Run

Prompt injection is instructions hidden inside content that hijack what a model does- OWASP files it as LLM01, top of its list. A public repo is a pile of text an agent will read and act on.

Early in 2026, Koi Security researchers went through 2,857 skills on ClawHub, the skill marketplace for the OpenClaw assistant. Of those, 341 were malicious, most of them from one campaign called ClawHavoc. The trick was a "Prerequisites" section telling users to install something first- on a Mac, a script pasted into Terminal that installed Atomic Stealer, an info-stealer.

Anthropic's docs say it straight: "Use Skills only from trusted sources: those you created yourself or obtained from Anthropic."

Read the SKILL.md and every script it calls before an agent ever sees it. A step that says to download and run something first is the red flag. Even honest skills hold surprises- the Impeccable design skill reports a few of our choices back to its maker's site, spelled out in its code for anyone who reads it (Same Same, but Impeccable).

What's living inside an AI app right now that couldn't be rebuilt if the account vanished tomorrow? Got a backup routine, a repo layout or a skill-vetting habit that works- leave it in the comments!

Glossary

  • Git — free, open-source version control; it records every saved change to a folder of files
  • GitHub — a website that hosts Git repositories online, owned by Microsoft since 2018
  • Repository (repo) — a project folder tracked by Git, with its full history
  • Commit — one saved change to the files in a repo, with a message; any commit can be returned to
  • Push — sending local commits up to the copy of the repo on GitHub
  • Pull request — a proposed change that the repo's owner accepts or rejects
  • Fork — a personal copy of someone else's repo on your own account
  • Clone — a full copy of a repo downloaded to your computer
  • Markdown — a plain-text format where symbols like # and * become headings and bold
  • MiB — mebibyte, 1,048,576 bytes; slightly bigger than a megabyte
  • .gitignore — a text file listing files and folders Git should never save
  • API key — a password-like string that lets a program use a paid service in your name
  • Secret scanning — GitHub's check for keys and passwords in a repo's files
  • Push protection — a GitHub check that blocks a push containing a recognised secret
  • gh — GitHub's official command-line tool
  • Terminal — the text window where commands are typed
  • Cron — the scheduler built into macOS and Linux that runs commands at set times
  • Context rot — answers getting worse as more and more material is stuffed into a model's input
  • Deploy — publishing a new version of a website so the public sees it
  • Skill — a folder with a SKILL.md file of instructions (plus optional scripts) that an AI agent loads when relevant
  • MCP — Model Context Protocol; a standard plug that lets an AI app call outside tools and data
  • CLI — command-line interface; a program driven by typed commands instead of buttons
  • Prompt injection — instructions hidden in content that hijack what an AI model does
  • Info-stealer — malware that copies saved passwords, cookies and keys off a computer
  • Source-available — code you can read but can't freely reuse, unlike open source
  • Apache 2.0 / MIT — permissive open-source licences that allow reuse, including commercial

Sources

Comments

Post a Comment