Has anyone tried using a review checklist during the AI writing phase itself?

Your ETL analogy is correct, but the overhead you're describing is exactly why we stopped calling it "generation" and started calling it "extraction."

We gave up on trying to make the LLM creatively follow a checklist. Now we treat our brand guide and tone rules as a source dataset. The prompt is just a query pulling specific, verifiable attributes from that dataset into the draft. It's less "apply a transformation" and more "SELECT active_voice FROM brand_rules."

The cost isn't in managing the pipeline, it's in maintaining the source data. If your brand guide is messy, your output is garbage, no matter how elegant your DAG is. That's the real parallel to data engineering.