AI Editorial Tools for a Publishing Platform

Any route that calls a paid model can be made expensive by someone who is not you. This one sits behind authentication, a sliding-window rate limit in Upstash Redis — ten requests per ten seconds, and five per fifteen minutes on authentication — and CSRF validation, with violations reported to Sentry rather than silently absorbed. The cost of getting this wrong is an invoice, not a bug report.

Each task asks for a length range: roughly 150 to 200 characters for an excerpt, 50 to 60 for an SEO title. A model treats that as guidance, not a contract, so the schema validates independently — the SEO title is rejected above 60 characters and the meta description above 160. The prompt shapes the output; the validation decides what is allowed to be saved.

The model provider sits behind a configuration module rather than being called from the route, with the provider selectable by environment. When one provider was dropped, that abstraction contained the call site — but not the migration: the dependency came out of the manifest and the old route was rewritten to return a gone response pointing at the working endpoint. An abstraction limits how far a change spreads; it does not make the change free.

Each task is scoped to a single field and works from content the editor has already written. The model is never asked to produce the article, and never decides that anything is finished. Generated text reaches a reader only after a person has read it, kept it and published it.