Work record / Stories

Graduated trust after a real mistake

I wanted AI help with LinkedIn comments, and built the system around an earlier failure.

I wanted automated help finding posts and drafting LinkedIn comments in my own voice. A real incident had already happened: comments written by AI went out publicly under my name with no review, and LinkedIn's own rules forbid automated account activity.

So the system did not get full autonomy from day one. I must approve each comment before it posts. Daily volume starts capped low. Hard stop rules end the entire run the instant the platform shows a warning or anything unusual, and it stays stopped until I say resume. It never has my password, it never solves a verification challenge, and it will not post, message or connect on its own.

The design aims directly at the earlier incident. AI comments under my name going out unreviewed is the exact failure I am guarding against. The practical risk is a platform restriction if the pattern looks automated, which is why the caps began low and the stop rules admit no exceptions.

My proposal builds graduated trust in from the start. The daily cap goes up only once I have gone three weeks in a row editing ten percent or fewer of the drafts I approved. My own formal go-ahead is still pending before anything runs for real.

When moving too fast without you has burned you, one slowdown is not enough. Put the slowdown into the design, so it cannot be skipped by accident. I designed around a real past mistake and did not pretend it never happened, and I kept myself as the approval gate on my own voice.

Story details

All stories · Work record