AI content review workflow
A workflow that reviews and scores content with an LLM before publishing.
- Category
- AI & RAG
- Tech stack
- Next.js
- MySQL
- Claude API
- Metrics
- LLM-scored entries with stored reasons
- Role
- Design + build
Content
An editorial workflow wired into the CMS: before content is published, an LLM pass scores it and records the reasoning, so quality issues surface in the admin instead of on the live site. The scores feed the content health views used across the admin.
Problem
Publishing without review risks uneven quality, but manually reviewing every entry does not scale.
Decisions
The score is advisory, not a gate. An LLM veto on publishing sounded like rigor and would have been the wrong kind — model judgments on niche technical writing are too noisy to hold final say, so a human still presses publish. What makes the score worth having is that the reasoning is stored with it: a low score with a stated cause is something you can act on, a bare number is a mood. The same scores feed the content-health views across the admin, so review effort compounds instead of evaporating after each publish.
Limitations
The reviewer scores one entry at a time. It has no view of the whole content graph, so it can't catch problems that only exist between items — near-duplicate coverage, for instance.
Outcome
Entries get an automated quality score with stored reasons before publishing.
How this was built
Built with AI assistance (Claude Code). The architecture, the decisions and the trade-offs are mine; a lot of the implementation was written in pair with the model. I work this way deliberately — it's why one person can run a system this size — and the reasoning behind every call here is documented above.