The process
A weak AI article is usually one prompt. This system uses separate passes so one fluent draft does not get to mark itself correct.
Research
Produces: Hard anchors and field 80/20 examples
Search for documented concentration in the field: named studies, datasets, standards, or honest adjacent-field hedges. Never invent a topic-specific 80/20 statistic.
Research brief
Produces: A narrow article claim, written down
Fill the brief before any outline: concentration claim, ignored majority, only-on-8020 idea, evidence tier, audience one-liner, and source candidates. The brief is public.
Draft
Produces: Article body
Write from the template spine and the brief. Topic-native voice. Name 80% and 20% in the intro. Start with prose, not a heading.
Fact-check
Produces: Citations, hedges, illustrative labels
Every exact percentage needs a named source or an honest hedge. Soft claims stay labeled. Invented scenarios get cut or labeled.
Voice pass
Produces: Broken symmetry and audience fit
Remove template cadence. Adjacent sections cannot end the same way. Fun topics sound fun. Serious topics sound warm and plain.
Second-model pass
Produces: A rewrite or review by a different model
A different model reviews or rewrites. It looks for hallucinated stats, generic how-to drift, and weak only-on-8020 blocks. One model does not grade its own work.
Publish gate
Produces: Checklist, credits, pass log
Run the publish checklist. Set authors, template, viz, SEO intent, and dates. Log the pass: date, pass name, model. If the gates fail, it is not ready.
Want to see this on a real article? Open the case study, or the "How this article was made" block at the bottom of any article.