We Didn't Automate Content. We Built a Marketing Teammate.

First thing we did was break apart what had been running on gut feel. On the left is where things used to be. Everyone had different standards, so corrections kept piling up: 'not on-brand enough,' 'too stiff,' 'cut the length.' By Draft v7 nobody was sure what the original even said anymore. On the right is what changed. We took the judgment criteria apart (audience, tone, forbidden phrasing, format, platform, purpose) and locked them into the system. The AI started working within those rules from the very first draft. Instead of corrections stacking up, the first draft already fit.

Those judgment criteria live in what we call the Ontology Decision Layer. Topic, audience, platform, purpose, tone, forbidden phrasing, all in one layer. On one side: topic discovery, context interpretation, and tone guardrails run in sequence. On the other side: Instagram carousels, Threads/X rewrites, and the publish queue. It's not a grand knowledge graph. It's a small concept system that a small team can start using right away. Once it was in place, the AI stopped behaving like a generic writing model. It started behaving more like a teammate working within a defined set of rules.

Draft generation came next. By the time a topic reached the model, it already carried its audience, angle, and format. The output stopped being 'plausible but off-brand' and started coming back with context baked in. That's where the downstream rework began to shrink.

The hardest part was making the writing sound human. Overblown claims, filler phrases, that flat AI cadence. If any of it stays in, people rewrite the whole thing regardless of how fast the draft came out. So we pulled tone review and humanizing into their own step. The system checks against the forbidden phrasing, allowed tone, and brand vocabulary defined in the ontology, and keeps rewriting until the draft lines up. A good chunk of what people used to do by hand now finishes here.

Rewriting the same piece for three platforms was always the dullest part of the job. It's not copy-paste. It's reshaping the format within the same set of rules. The original draft goes through rewrite rules (length adjustment, tone alignment, CTA repositioning) and comes out as versions for Instagram, Threads, and X. One topic pass, three channel versions. If a version blows past the platform limit, it loops back and rewrites itself.

From the operator's side, it all came down to one question: can I see the whole day on one screen? Topics in the inbox, draft status, tone review pass/fail, card news check, publish queue. We simplified the operations view into a single pipeline so that no new manual work piles up after the draft stage. It stopped feeling like a tool that writes for you. It started feeling like handing over the ops flow itself.

Content doesn't end with text. You need card news slides, and each slide needs a supporting image. That's where the designer bottleneck used to be worst. Now, once a draft passes review, it splits into slides, each slide gets an auto-generated illustration, and the whole set renders as 1080x1350 card news images on an HTML template. 'Copy is done but waiting on images' stopped being a thing.

Last step: publishing. Threads and Instagram run as one sequential flow. Two or more images and Instagram switches to carousel automatically. If the platform rejects an image upload, there's a JPG conversion retry and a hosting fallback chain so one small failure doesn't stop the whole run. We also added a pre-run environment check, so troubleshooting moved from 'fix it after it breaks' to 'catch it before it runs.'