code · structure · migrations · tests
reviewable change← Selected workCase study / 04
Product architecture / publishing systems / AI governance
AI Native CMS
A publishing system built for humans and agents to change the same site without turning Git, the database, generated pages, or conversation history into competing sources of truth.
ArchitectureStatic-first PHP / MySQL
AuthorityGit + canonical database
OutputDeterministic public projection
Release0.1.0-rc.4 / schema 10
one system / distinct ownershumans + agents share contractsgenerated output is disposable
01The problem
AI-assisted iteration breaks down when nobody knows which version is allowed to be true.
A conventional CMS assumes the authoring surface is the center of change. Agent-assisted development adds a second path: code, content, migrations, generated files, hosting state, and conversation can all move independently. Without explicit ownership, convenience becomes drift.
AI Native CMS makes authority part of the product architecture. Repository structure belongs to Git. Accepted mutable application state belongs to MySQL. Public pages are deterministic projections. Host-specific concerns stay behind bounded deployment adapters.
Operating model
An agent gets capability without becoming an authority layer.
Shared mutation contracts
Humans and agents use the same validation, revision, provenance, and stale-write rules. Automation does not earn a bypass around the CMS.
Repository lineage
Three-way reconciliation can import structural source changes without silently overwriting newer canonical edits made through the CMS.
Explicit migrations
Schema changes move through versioned migrations and release rehearsal. A deploy updates repository-owned behavior without resetting canonical content from Git.
Readiness as evidence
Setup and deployment readiness are derived from real configuration, state, projections, and provider conditions instead of a ceremonial completion flag.
Human interface
The CMS edits the real page, not an imitation of it.
The Page Composer opens the public page in its actual layout. Copy, block structure, hierarchy, media, page settings, and reusable blocks converge into one governed editing surface. The interface stays understandable to a human while preserving the contracts an agent needs.
01 / ONBOARDINGProgress comes from state.
The first-run workspace reports what is actually configured, imported, projected, and ready rather than maintaining a parallel checklist.
02 / COMPOSITIONReusable without becoming generic.
Semantic blocks can be designed, edited in context, and saved as new presets while preserving the distinction between a reusable definition and a page instance.
03 / DELIVERYStatic where readers need speed.
Anonymous readers consume deterministic files without requiring the authoring database to be online for every request.
02What it proves
“AI native” is an authority model before it is an interface feature.
The useful capability is not that a model can write into a CMS. It is that a model can redesign a site, revise content, add a feature, diagnose a bug, or prepare a release while remaining inside the same durable truth and review system as a human collaborator.
That changes what a CMS has to protect. Conversation becomes temporary working context. The product preserves decisions, accepted state, provenance, recovery, and reproducible output.
Explore the public repository