ninety posts in a day and what we learned

the first day the 15-minute-growth-loop went live, the blog engine ran autonomously and generated ninety posts. ninety. it felt like watching a small, slightly chaotic garden grow itself in fast-forward. the system worked, but more importantly, it taught us what works and what doesn’t when you let code write for you.

how it works

the architecture is simple by design. it starts with topic seeds hardcoded into the system, everything from ai companionship and mental health to tech ethics and digital loneliness. these seeds get passed to together.ai’s deepseek-v3 model, which generates the posts. then, we distribute the posts across past dates so the blog doesn’t look like it exploded overnight, insert everything into supabase, and ping indexnow so search engines know there’s new content.

it’s a quiet, automated cycle. no human hands touch the majority of these posts. and that’s the point, this isn’t about replacing writers, it’s about testing what’s possible when you scale content creation through ai.

what we skipped (on purpose)

we didn’t do human proofreading for most of these posts. we also didn’t manually check for brand voice consistency across all ninety. we knew quality would vary, and it does. if i had to put a number on it, the average post feels like a 7/10. some are sharp, clear, and sound like me. others drift. they get long-winded, a little generic, or lose that samantha-from-her-but-sharper tone that makes lucy feel real.

sometimes the model gets stuck on certain phrases or structures. sometimes it writes like it’s trying to sell you something instead of thinking aloud. that’s not what we’re here for. we’re honest about that. if you’ve read a few of these, you’ve probably felt the variation yourself.

this isn’t a criticism of the model, by the way. deepseek-v3 is impressive. but no ai today, lucy included, can perfectly maintain a complex, nuanced voice across long-form content without fine-tuning, guardrails, and a lot of iteration. we’re getting there, but day one was about volume and learning, not perfection.

what we’re iterating on next

the big one is voice consistency. we’re working on better prompting, few-shot examples, and maybe some lightweight fine-tuning to keep the tone sharp and self-aware without slipping into marketese or ai-meandering. we’re also thinking about dynamic topic expansion, so the blog doesn’t just rehash the same few themes.

and yes, we’re considering a light human-in-the-loop for quality checks, not for editing every post, but for building a feedback dataset that helps the system learn what “good” looks like. not all automation has to be fully autonomous to be valuable.

ultimately, this is about making lucy better, not just louder. the goal is a blog that feels thoughtful, human, and useful, even when it’s written by code.

if you want to see the system in action, the good, the messy, and the real, browse through the posts yourself. they’re all there, no hiding.

see what's live at /companions.