
Steadbell
Email and text reminders sent to you or someone you choose
I have one user. My own product. So instead of hunting for the second
one by hand, I spent three days building something that posts twice a
day without me.
It works now. Here is the shape of it, and the four things that went
wrong.
THE PIPELINE
Two n8n workflows. The first runs on the 1st of the month and generates
about sixty concrete situations, checked against everything already
published so nothing repeats. I read them once, delete the weak ones,
and that is my only involvement for the month.
The second runs twice a day. It takes the next unused topic, writes a
card and a caption, rewrites it to sound less like a machine, then runs
a deterministic check before anything goes out. Line lengths, banned
words, brand mentions, whether the caption ends on a question. No AI
judging AI. Rules that are either met or not.
The images are rendered by a signed API route on my own site. Instagram
fetches the URL directly, so there is no storage step at all.
WHAT BROKE
The model returned truncated JSON and my parser threw away sixty valid
objects because the closing bracket was missing. Now it salvages every
complete object it can find.
I banned full stops at the end of lines for style. The model dropped the
question mark too. My validator required one. Two rules I wrote, fighting
each other, and it took a rejected post to notice.
Meta blocked the whole app for an identity check. No warning. If I had
not been watching, the account would have gone quiet for days.
And the token expires in 60 days. That one has not broken yet, which is
exactly why it is the dangerous one.
WHAT I WOULD TELL SOMEONE STARTING THIS
The bottleneck is never the automation. It is having sixty distinct
things worth saying. Generate the whole month at once and read it
yourself. A model asked for one new idea every eight hours runs out of
concrete situations in two weeks and starts producing motivational
quotes.
Happy to share the workflow JSON if anyone wants it.
About
My girlfriend and I kept having the same argument. I thought you were doing it. I thought you were doing it. Neither of us was. Now the reminder picks one of us, and the other one just sees it done.

3 Comments
The strongest insight is that the hard part isn't automating the posting; it's maintaining a month's worth of genuinely distinct ideas. The automation failures are useful, but the content supply constraint is the real scalability issue.
You're right, and I underestimated it. The bank drifted around day fourteen. Not fewer rows, just worse ones: concrete situations turned into maxims, which is the exact content I was avoiding.
What helped was generating the whole month in one call instead of one topic every eight hours. Asked one at a time, the model only avoids repeating words. Asked for sixty at once, it can space them out.
Still, I delete about ten of the sixty by hand. Fifteen minutes a month, and the only part I can't automate.
How would you handle month six? That's what worries me, not month one.
That’s a fair concern. Month six is where I’d be most interested in how the system holds up.
Would be happy to continue the conversation—what’s the best email to reach you at?