GitHub for marketing teams: why you will be there within a year
Once agents write your drafts, your brand documents stop being reference material and start being runtime configuration.
A plain-text version of this site for AI assistants is at https://contengi.com/llms.txt, with the full product documentation in one file at https://contengi.com/llms-full.txt. Every page is also available as markdown by adding .md to its URL.
TL;DR
- Marketing teams end up on GitHub because the inputs to an AI content operation are plain text files, and a repository is the cheapest place to keep text that several people and several agents all change.
- Claude Code reads a project CLAUDE.md at the start of every session, so your tone of voice stops being a document nobody opens and becomes something an agent applies without being asked.
- A CODEOWNERS file automatically requests review from the people who own a folder whenever a pull request touches it, which gives you sign-off without buying a workflow tool.
- GitHub Agentic Workflows run natural-language instructions as GitHub Actions, inside firewalled containers with read-only tokens by default and write access limited to outputs you declare.
- Push a content branch to a Vercel-connected repository and the pull request gets its own preview URL, so a reviewer reads the finished page instead of a document that looks nothing like it.
Why a marketing team ends up on GitHub
Three years ago, telling a developer your marketing team worked in the repo would have been a decent joke. In 2026 it reads like a calendar invite.
The reason is dull, which is usually how these things go. Once agents write your drafts, your brand documents stop being reference material and start being runtime configuration. The tone of voice file is read before every draft. The playbook is read before every brief. The positioning line in your strategy doc shapes 40 blog posts a quarter, so you start caring quite a lot about which version of it was in play last Tuesday, who changed it, and whether anyone reviewed the change.
Git was built for exactly that question. Every change is a commit with an author, a timestamp, and a readable diff. You can see that someone softened the ICP description on 14 March, see what it said before, and put it back in eight seconds.
Claude Code makes this concrete. It reads persistent project instructions from a CLAUDE.md or AGENTS.md file at the start of every session, and a project-level file is shared with the whole team through source control. Keep it under 200 lines, because longer instruction files consume more context and reduce how consistently they are followed. One thing to know before you paste your entire brand book in there: these files are treated as context rather than enforced configuration, so they guide behaviour instead of guaranteeing it.
That single detail is what pulls marketing into the repo. The brand rules have to sit next to the work for an agent to read them, and the work is text.
We run Contengi this way. The brand rules, the anti-slop standard, and every blog playbook are files an agent reads before it writes a word, and version history is the only reason I can tell you exactly when a rule changed and what it replaced. After a decade of watching brand guidelines rot in a shared drive, that part still feels like cheating.
What goes into a marketing repository
Brand rules, playbooks, briefs, research notes, and website copy in markdown. Campaign videos, raw design files, and live dashboards stay where they are.
A working marketing repo looks like a folder tree a new starter could navigate on day one: a brand/ folder with positioning, tone of voice, hard rules, and ICP profiles, a playbooks/ folder with one file per content format, a research/ folder for transcripts and customer calls, and a content/ folder for posts if your site reads markdown. Nothing clever. The cleverness sits in what reads those files.
Skills are the next layer. A SKILL.md file with YAML frontmatter and markdown instructions turns a procedure into something an agent runs, and the body only loads when the skill is used, so a long reference document costs almost nothing until it is needed. Commit the folder and everyone working in that repository has it. The format follows the Agent Skills open standard, which works across multiple AI tools, so your editorial process does not get locked inside one vendor's box.
Our own AI slop cleaner ships as exactly this, a skill you unzip into a skills folder, which then strips the hidden characters and rewrites the sentences that read like a machine wrote them. A repo full of files like that is a content operation anyone on the team can run.
The research/ folder earns its place faster than the others. Sales call recordings, webinar transcripts, and support tickets are the only material in your whole operation that nobody else has, and dropped into a folder an agent reads, they turn a generic draft into something with a customer's own words in it. Transcripts are messy, long, and perfect for this, because an agent happily reads 12,000 words of rambling to find the one line a prospect used about pricing.
Agencies get a second benefit here. One repository per client, each with its own brand folder and its own owners, means the same playbooks run across every account without the voices bleeding into each other. Fork the structure once and onboarding a brand becomes a morning's work instead of a fortnight of back and forth.
The rule of thumb for what goes in: if an agent needs to read it to do the job properly, it belongs in the repo.
| Part of the operation | Where it belongs | Why |
|---|---|---|
| Tone of voice, hard rules, positioning | Repository | Read before every draft, so version history is the whole point |
| Playbooks and editorial SOPs | Repository | A procedure an agent follows, better as a skill than a wiki page |
| Briefs and research notes | Repository | Cheap to store, useful as context, and diffable |
| Blog posts and site copy | Repository, if the site reads markdown | Content ships with the deploy and gets a preview URL |
| Blog posts on WordPress or Webflow | Stays in the CMS | The database is the source of truth, the repo holds process |
| Campaign video, raw design files, big image sets | Stays out | Binaries bloat a repo and the diffs tell you nothing |
| Dashboards and live reporting | Stays out | Point at the data from the repo, do not copy it in |
| Social scheduling and calendar views | Stays out | Git has no calendar, and pretending otherwise ends badly |
How review works when content lives in a repository
Through pull requests, with named owners per folder. The mechanism is a CODEOWNERS file, which automatically requests review from the individuals or teams responsible for the files a pull request changes. Put legal on /content/claims/, put your brand lead on /brand/, and the right person gets pinged by the system rather than by someone remembering to ping them.
A few details save you a frustrating afternoon. The file goes in the .github/, root, or docs/ directory, and where more than one exists, GitHub searches those locations in that order and uses the first one it finds. Order inside the file counts too, because the last matching pattern takes precedence, so your catch-all line goes at the top and your specific folders below it. Owners need write access to the repository, teams included, even when every member already has it individually.
Then make it binding. Turning on the branch protection option to require review from code owners means a content change cannot merge until an owner approves, and an approval from any one of the listed owners satisfies the requirement. Code owners are not requested on draft pull requests either, which gives a writer somewhere to work in the open without summoning four reviewers to a half-finished post.
This is the part that converts the sceptics. A comment thread on a Google doc is a conversation. A pull request is a decision with a record attached, and six months later you can see who approved the pricing claim.
Reviewing the page rather than the draft
Connect the repository to Vercel and the review gets better again. Vercel deploys every push by default, including pushes and pull requests made to branches, and the latest push to any pull request lands on a unique preview URL that arrives as a comment on the pull request itself. Your reviewer clicks a link and reads the finished page, with the typography, the hero image, and the internal links all in place.
Rollback comes free. Revert a commit that already went to production and the previous production deployment becomes available at the custom domain instantly. Anyone who has ever pushed a wrong price to a live pricing page will understand why that is worth the setup time.
If the site is headless, this is where markdown in the repo stops being a developer preference and becomes a content layer your whole team works in. A site Claude Code built will often read MDX files straight out of a folder, with no admin panel anywhere.
What a repository does that a shared drive cannot
It runs things on its own. This is the newest piece of all of this, and the one worth clearing an afternoon for.
GitHub Agentic Workflows are AI-powered repository automations defined in markdown and run as GitHub Actions workflows. Each file has two parts, YAML frontmatter that sets the triggers, the permissions, and which write operations are allowed, and a markdown body holding the instructions in plain language. You compile the markdown into a .lock.yml Actions file, and from then on the repository reasons about work instead of following fixed if-then rules.
The guardrails are the interesting bit for anyone who has to answer to a security team. Agents run in firewalled containers with read-only tokens by default, write actions are limited to the safe outputs you declare, and those are checked by agentic threat detection. You need Actions enabled, an account with an AI engine such as GitHub Copilot, Anthropic Claude, OpenAI Codex, or Google Gemini, and the GitHub CLI installed.
The documented examples are developer jobs, triaging issues, investigating failed builds, generating weekly status reports, and keeping documentation in step with the code. Translate them and you get the marketing versions:
- A weekly performance digest. A scheduled workflow reads your search data, opens an issue listing the pages that slipped, and tags the owner of each one.
- A refresh queue. The same job flags posts whose impressions are climbing while clicks fall, then drafts the brief for the rewrite.
- A brand check on every content pull request. The workflow reads the diff against your hard rules file and comments where a draft broke one, before a human reads a word.
- Documentation that keeps up. When the product ships a change, the workflow opens a pull request updating the pages that describe it.
That last one is where the money is. Content going stale slowly is the most expensive thing in any content operation, and a repository notices the change the moment it happens.
One trap, since nobody mentions it until it bites. If you filter a workflow by path so it only runs when files under /content change, a workflow skipped by path filtering leaves its associated checks in a pending state, and a pull request that requires those checks is blocked from merging. Your brand-check workflow will happily block every design-only pull request in the repo until you work out why.
Do marketers need to learn git?
No, and anyone telling you otherwise is selling a course. There's three ways in, and only one of them involves a terminal.
Edit in the browser. GitHub's web interface lets you create, edit, and delete files, write a commit message, and open a pull request without installing anything. For a writer changing a markdown post, this is the whole job.
Use a desktop client. GitHub Desktop syncs a folder on your machine with the repository, so your writers work in whatever editor they like and press a button marked commit. The folder behaves like Dropbox with a memory.
Let an agent drive. This is where it gets good. Claude Code handles the git workflow for you from the terminal or your IDE, so a content manager describes the change in English and reviews the diff rather than typing git rebase and hoping.
Be honest about what will go wrong, though. In the first month someone will commit straight to main because the branch rules were never turned on, two writers will edit the same paragraph and hit a merge conflict neither of them can resolve, review will turn into a queue because every typo fix needs an approval, and somebody will commit a 200 MB PDF. Branch protection fixes the first, smaller pull requests fix the second, and the queue is a policy question, so decide early which folders need an owner and leave the rest open.
Where GitHub is the wrong answer
Plenty of places. Your designers live in Figma and should stay there. Video belongs in a storage bucket, because committing a 400 MB file to a repository helps nobody and git will remember it forever. Your CFO is not going to review quarterly messaging in a diff view, so send a preview URL and let the comment come back however they want to send it.
If your site runs on WordPress, Shopify, Webflow, or Framer, the posts live in a database and the repository holds your context and your process. That split works fine, and keeping the knowledge base somewhere every agent reads first is the part that changes output quality, regardless of where the published page ends up.
A repository also gives you no campaign calendar, no social scheduler, and nothing resembling a Gantt chart. GitHub Projects covers tracking well enough, it will not replace the tool your team already uses to plan a launch.
How to set up a marketing repository in a week
Day one: create a private repository and name it for the operation, not the team. Something like brand-ops ages better than marketing-stuff. Add four folders, brand/, playbooks/, research/, and content/, and nothing else.
Day two: write the brand files as rules rather than prose. Explicit instructions land better than adjectives, so write the tone of voice section with examples of what right and wrong sound like. Hard rules in a numbered list. Positioning in a paragraph a stranger could repeat back to you.
Day three: add a CLAUDE.md that points at the rest. Keep it under 200 lines, and let it say where things live and which rules are never negotiable. Every session reads this file, so it earns the editing time.
Day four: turn one repeated procedure into a skill. Pick the thing you explain to a new writer every time, your blog brief process or your refresh checklist, and write it as a SKILL.md. Commit it, and it belongs to the team rather than to whoever wrote it.
Day five: add CODEOWNERS and turn on branch protection. Two or three lines to start, a catch-all owner at the top and your sensitive folders below. Require code owner review on the main branch and nothing merges unreviewed.
Week two, once it sticks: connect the site and add one workflow. Wire the repository to your host so every pull request gets a preview URL, then write one agentic workflow and only one. The weekly performance digest is the easiest to trust, because its output is a list of pages and the worst case is a boring issue nobody acts on.
Then check what the bots see. Content in a repository still has to render in the HTML a crawler downloads, so run the pages through the AI bot view before assuming an AI engine can read them.
The whole setup is a day of real work spread over a week of calendar time. What you get at the end is a content operation where the rules are versioned, the approvals are recorded, the preview is the published page, and the recurring jobs run whether or not anyone remembered them.
That is what an agentic content workflow with approval gates looks like when you build it yourself. Some teams would rather it arrived already built, which is a separate conversation and a fair one.
Frequently asked questions
Why would a marketing team use GitHub?
Because the inputs to an AI content operation are text files, and GitHub versions text files better than anything else. Brand rules, tone of voice, playbooks, and briefs become configuration that agents read before every draft, so you need to know which version was live, who changed it, and who approved the change.
Do you need to know how to code to use GitHub for marketing?
No. You can create and edit files entirely in GitHub's web interface, sync a folder with GitHub Desktop, or describe the change to an agent that handles the git commands for you. The skill you need is writing clear rules, which is a marketing skill already.
How do content approvals work on GitHub?
Through pull requests and a CODEOWNERS file, which automatically requests review from the owners of the files a pull request changes. Enable the branch protection option requiring code owner review and the change cannot merge until an owner approves it, with one approval from any listed owner being enough.
Can AI agents run marketing tasks inside a GitHub repository?
Yes. GitHub Agentic Workflows let you write instructions in markdown with YAML frontmatter, compile them into GitHub Actions workflows, and run them on repository events or a schedule. Agents run with read-only tokens by default and write only through the safe outputs you declare.
Should your blog posts live in a repository?
Only when your site reads markdown, as a Next.js or Astro build usually does. On WordPress, Shopify, Webflow, or Framer the posts stay in the CMS database, and the repository holds your brand context, your playbooks, and your process instead.