> A markdown-native, git-based website that Claude Code or Cursor can edit directly, with schemas and CI checks instead of a CMS panel.

Canonical: https://augustynglowacki.com/services/ai-editable-no-cms/

Schema discipline, not vibes.

# AI-editable websites, no CMS

This is a website with no CMS: content lives in the repository as structured markdown or YAML, validated against a schema at build time, and edited directly by an AI coding agent under a pull request workflow. There's no admin panel to log into. The guardrails that would normally live in an editor's UI - required fields, valid references, section limits - live in code instead, checked by CI before anything ships. This site is built that way; see [how this site is edited](/how-this-site-is-edited/) and [projects built the same way](/projects/).

[Talk about your site](#lead-form)

## How it works, who it fits

### What is this, concretely?

Astro content collections (or the equivalent in your framework) with strict schemas, one file per page per locale, sections defined as a typed union so an invalid section fails the build instead of shipping broken. Content changes go through Claude Code or Cursor, get reviewed as a diff, and merge like any other code change. Lee Robinson, ex-Vercel, migrated cursor.com off a headless CMS to raw code and markdown in three days for $260 in agent tokens: "With AI and coding agents, the cost of an abstraction has never been higher... do we really need a CMS?" ([leerob.com/agents](https://leerob.com/agents)). **That's the bet this service is built on.**

### Who does this fit?

Teams where a developer or a technically comfortable founder is already the one touching content, where pages change in shape as often as they change in words, and where the AI tooling should be able to add a new section type as easily as it edits a paragraph. **It fits well alongside the audit and rescue services**, because a markdown-native site is trivially compliant with the technical checks both of those cover.

### Who does this not fit?

Teams where non-technical editors need to build and rearrange sections themselves, on their own schedule, without a pull request in the loop. **If your marketing team ships campaign pages weekly and nobody on it reads a diff, that's a CMS case** - see [AI-editable websites with a CMS](/services/ai-editable-website-with-cms/). The counter-position is worth stating plainly: markdown-in-git can reintroduce CMS complexity over time as assets, permissions, localization and scheduling needs grow ([Knut Melvaer, summarized by Graeme Fulton](https://prototypr.io/post/cms-markdown-ai-agent)). I scope for where that line sits before I build.

### What guardrails ship with it?

Five things, every time: Zod (or equivalent) schemas on every content type with descriptions on every field; **CI validation that fails the build on an invalid section, a broken reference, or a missing required field**; an AGENTS.md at the repo root written for the coding agent as its actual user; deploy previews on every pull request; and markdown delivery for every page - a .md twin at the same URL plus Accept: text/markdown negotiation, which costs nothing extra once content already lives as markdown. I don't sell that overlap as a citation lever; see the [llms.txt post](/posts/llms-txt-will-not-get-you-cited/) for why. Without that discipline, the same gap that costs your editors hours hand-propagating one edit is invisible to AI engines too: no entities, no consistent facts to extract, contradictory copies of the same claim across pages.

### What does the market actually think of this?

Split, and I say so. TinaCMS's own homepage now leads with "Markdown is the Answer - LLMs understand Markdown natively" and rates itself "GEO readiness: excellent" - **a CMS vendor conceding the premise**. GitCMS markets itself explicitly as turning "ChatGPT or Claude into Content Agents." The skeptical case is real too: complexity that a CMS handles for free - asset pipelines, granular permissions, multi-locale scheduling - has to be built by hand in a git-native setup, and it shows up later rather than never. I scope the engagement against which side of that line your team sits on, not against which side sounds better in a pitch.

### What if the scope is bigger than one engineer?

Then I say so early. One engineer is the right shape for a marketing site, a rebuild, or a rescue - not for a platform that needs several people shipping in parallel to hit a fixed date. I work at [Bejamas](https://bejamas.com/) day to day, so for that kind of scope I will introduce you there instead of stretching a solo engagement to cover it. **You get the capacity the work actually needs, and I would rather hand the project on than miss your deadline.**

## The process

### You send the brief

What the site needs to do and who's actually going to edit it - a developer or a technically comfortable founder. One paragraph is enough to start.

### You get a fixed number in writing

A scoped price for the schemas, the CI checks, and the AGENTS.md, starting at 3,500 USD - set before work starts.

### Work ships on that number

Content, schema, and CI validation land together, reviewed as a pull request like any other code change.

## See if this fits your team

Contact form on this page: name, email, and what you need. [Talk about your site](#lead-form).