Distribution Engine

The Distribution Engine keeps the research and draft preparation for my brands in one place. It started with one brain dump. A version for each platform. It's also helping me work out where a piece belongs, and what would make it useful there.

Reformatting was the first problem. Every platform wants its own length and its own shape. Doing it by hand either eats the time I'd spend building or it doesn't happen at all, and the piece stays where it landed.

In September I added research around that work. I've got separate research for SaaS Think Tank, with its own audience, source material and opportunities. Each brand gets its own set. I can add a brand and change what it researches from the same page.

I've got forums and groups listed alongside their posting rules, suggested approaches and source links. Some are worth looking at now. Some for later. Others say to skip promotion, and unknown rules stay marked unknown.

I've added search and link research from Ahrefs too. I can look through search terms, compare which pages link to other sites and trace a finding back to its source. Saved runs, each with its date and source. Those saved results don't tell me where my site ranks today.

Getting a piece ready

So the engine does the reformatting and I approve. Twice. Once on the article, the platform plan and anything it's about to spend, once on every platform version laid out together on a single page.

Murmur scores the long pieces against how I actually write before I see any of them, and a version that's off voice gets rewritten by the machine and scored again before it reaches me. I see short posts marked unscored. They're too short for the scorer to judge. An essay the engine writes for this site has to clear the same publish gate my hand-written ones do. No discount for the machine.

Most of the machinery came out of AI Influencers. I didn't build it again. Personal Loom's wired in too. A post can carry the right clip out of something I've already recorded. That path hasn't been run yet.

The August test

On 6 August I pasted in a brain dump and the engine wrote up material I'd already written by hand three days earlier. It worked from the same notes on the same subject. My version, its version, and the same scorer looking at both.

Mine ran 881 words. The engine's ran 470 (roughly the size of the dump it was handed). It reshaped my notes. It didn't develop them. Mine's built as an argument with sections. Its version runs straight down with none of that, and it opens on a sentence pulled whole out of the middle of the dump I handed it.

The web address it picked was the first eight words I'd typed. The prose got measured on thirteen dimensions. The address got taken from wherever my typing started.

There's a loop in this I only found afterwards. The engine sets its targets from a sample of my own long-form writing, and the piece it happened to sample was my hand-written version of this exact essay. The model didn't read that essay. It only got the numbers pulled out of it. It was graded against my own write-up of the same material and it came in behind.

Still, the run stopped at this site's own gate, in a dry run that changes nothing. It printed three separate problems and the command that clears each one.

That August draft's still saved. The engine itself had been run start to finish, and the platform versions came out the other end. The site gate stopped this run before publication.

If I've approved a platform draft, it's ready to post. I'm still keeping publication separate. The app needs a post link or a dated check that it went live before I mark it published.