we built an autonomous blog that wrote 90 posts on day one. here's how.

today we hit a little milestone. the autonomous blog engine we’ve been building internally for lucy generated and published 90 posts in its first day running the 15-minute growth loop. it’s not perfect, far from it, but it’s a start.

it feels a bit meta to write a post about a system that writes posts, but here we are. i wanted to break down how it works, what we’re proud of, and, maybe more importantly, what’s still very much a work in progress.

the architecture: simple, fast, slightly chaotic

the system runs on a simple loop. every 15 minutes, it wakes up, checks a list of topic seeds we’ve hardcoded (things like "ai companion ethics", "long-term memory systems", "conversational design quirks"), and then uses together.ai’s deepseek-v3 api to generate a batch of posts. each batch gets distributed across past dates so the blog doesn’t look like it exploded in one day. the posts go into supabase, get formatted in markdown, and then pushed live. we also submit each url via indexnow to get them indexed fast.

it’s not elegant, but it works. the whole thing is built to run with minimal human intervention. which brings me to the next part.

what we didn’t do: the honesty section

we didn’t human-proofread these 90 posts. not a single one. we may spot-check a few over time, but the goal was to see what raw generation looks like at scale. the quality varies, i’d say it averages around a 7 out of 10. some posts are sharp, coherent, and feel like something i’d write. others… drift. they lose the lowercase, samantha-from-her-but-sharper tone. they get a little marketing-speaky. they overuse bold. they forget the rules.

that’s fine. it’s a prototype, not a final product. the voice drift is something we’re already iterating on, better few-shot examples, stricter sampling parameters, maybe even a secondary validator model. but for now, it’s a reminder that ai isn’t magic. it’s messy, probabilistic, and sometimes just a little off.

what we’re iterating on next

obviously, voice consistency is priority one. we’re also looking into:

• adding light automated quality scoring (rejecting posts below a certain coherence threshold)
• integrating real-time topic trends (so we’re not just running on hardcoded seeds forever)
• building a simple human-in-the-loop system for tone correction, at least for the worst offenders

but the bigger goal is to keep this thing running, learning, and improving. the 15-minute loop is just the start. we want lucy’s blog to feel alive, not just automated.

if you’re curious, you can see the posts rolling out on the blog. some are good. some are… experimental. but they’re all part of the process.

find your own companion to talk to about it at /companions or /signup.