For about two months I shipped things and wrote about none of them. Orbit got a payments engine, Pulso became a native iOS app, the website got a real intake form. Writing it up was always the last item on the list, which is another way of saying it was never the item.
So I built Pluma, an AI content pipeline that reads my project tracker and drafts the posts for me.
The shape is four agents running in sequence. Scout queries Orbit for tasks I completed in the last seven days and throws out the busywork. Strategist groups what survives into themes and picks the platform mix. Writer drafts in my voice. Editor does a platform fit and SEO pass, then drops everything into a review queue. Nothing publishes on its own, X and LinkedIn stay copy and paste by me, and a blog post only goes live when I approve it.
The interesting part was not the automation. It was the arguments.
The 94% that meant nothing. The first design put an AI confidence badge on every draft, a number in the low nineties, glowing nicely. I asked what computed it. Nothing computed it. It was decoration shaped like data. It got stripped out and replaced with two scores actually derived from the row: a readiness score built from whether a draft has the fields its platform needs, and a style compliance score checked against my own written voice guide. Lower numbers, real ones. It was a beautiful badge telling me a comfortable lie.
The voice had to live outside the code. The Voice tab holds my style guide in four sections: who I am, how I sound, how I structure a piece, and how I frame the work. The scheduled agent reads it live before drafting anything. When my writing changes I edit a text box, not a prompt buried in a repo, and the next Monday run picks it up with no deploy.
Show the reasoning or do not bother. Every draft now stores why its theme was chosen, and blog drafts store the SEO decisions as notes alongside the copy. If I cannot see the editorial judgment I cannot correct it, and a pipeline I cannot correct is one I quietly stop trusting in about three weeks.
The backlog test. Before trusting the weekly cadence I ran a one time catch up across roughly two months of finished work. It produced 24 drafts. That was the point it stopped feeling like a toy, because the raw material had been sitting in the tracker the whole time, already written down, just never turned into anything.
What is still unsolved: the voice is close but not mine yet. There is an approve-driven feedback loop wired in, so approved drafts become the reference set the next run learns from, but with zero approvals so far it has nothing to learn from. I would rather leave that honestly unsolved than claim a fix I have not proven.
If you are building something similar, the transferable lesson is not the agent count. It is that every number your interface shows should trace back to a field in your database, and if it does not, delete it. A fabricated confidence score is worse than no score at all, because it trains you to trust the wrong thing, and it does it politely.
