The problems people arrive with

Managing website copy stored in TypeScript

The most common situation and the one the product was designed around.

The problem

Copy lives in exported objects that drive routes, metadata and navigation. Changing a word is a pull request, a review and a deploy — so small improvements never happen and the page drifts out of date instead.

What helps first

An inventory. Before anything else, knowing what is in there — how many headlines, how many CTAs, how many near-duplicate paragraphs — changes the conversation from "our content is a mess" to a list with a size.

What comes later

Editing those values from outside the codebase, with the change arriving as a pull request your developers review exactly as they would any other.

Letting marketing edit a Next.js site without giving up Git

Why the usual answers disappoint

A headless CMS means migrating the content model and adding a runtime dependency. A visual builder means rebuilding the site around its components. Both are large projects justified by small edits.

The Git-connected alternative

Leave the content where it is. Index it, give every value an identity and an exact source address, and make approved edits arrive as pull requests through the pipeline that already exists.

What developers keep

Source control, code review, CI, deployment and the ability to revert. Nothing new sits on the rendering path.

Auditing content across multiple repositories

Why it is hard today

Tools are scoped to one project. Duplication and contradiction across a portfolio are invisible to all of them, which is precisely where the expensive inconsistencies live.

What the graph adds

One index across every connected repository and brand, so "where does the group claim this?" is one query rather than nine separate audits.

Where to start

Connect two repositories that should agree with each other. The first duplicate-messaging report is usually persuasive on its own.

Stopping product claims drifting from product reality

The mechanism that works

Separate the fact from the copy. Give the fact an owner, evidence and a validity window. Link every claim to the fact behind it. When the fact changes, the dependent content becomes a list.

The specific failure to catch

A planned capability described in the present tense. It is the most common and most damaging content defect a software company produces, and it is mechanically detectable once availability is a fact rather than an adjective.

An honest caveat

Detection is only as good as the facts behind it. A claim with no linked fact is findable immediately; a claim whose fact silently went stale needs somebody to have maintained the validity window.

Structured content for AI agents

What agents actually need

Not more prose. Addressable content with types, clear entities, stated relationships and evidence for claims — so retrieval returns the right twelve items rather than a page of approximations.

Why the content model matters more than the prompt

A model given the company's real facts, current claims and audience definitions produces usefully different output from the same model given a longer instruction. The context is the variable.

The by-product

Content structured well enough for your own agents is also structured well enough for AI search. That is a consequence of the model, not a separate feature to buy.