If you're building any kind of SNS automation tool, the fight against bot detection is part of the job. This post is the postmortem of the day Meta's automation detection caught one of the Instagram accounts I was running through my indie SaaS (GramShift), what the actual trigger pattern looked like from the operator side, and the design rewrite I shipped in v1.5.0 as a response.
Note up front: the specific thresholds, probabilities, and timing distributions inside the pacing engine are the product's moat and are intentionally not in this post. The shape of the rewrite is what's interesting; the exact numbers stay private.
What happened
I detect the event from my own logs first, not from Instagram. The like count for a running cycle drops far below the expected band, and the operation log starts emitting action_blocked. I open the affected account in a browser, log in, and the dashboard greets me with:
We're sorry, but we limit how often certain actions can be performed.
