We've been publishing 2 to 4 short-form reels a day on TikTok, Instagram and YouTube since Oct 30. The screenshot is our real dashboard for the last 30 days: 95 posts, 39.2K views, 2.1K saves. Most reels did okay. One clearly beat the rest: 3,879 views and 209 saves on Instagram, 2,302 views and 138 saves on TikTok, and 70 people commented a keyword on it. What it did differently: 1. The hook put a free option right next to a paid tool. People stopped because they thought they might be overpaying. 2. It was useful enough to come back to. Saves pushed reach more than likes did. 3. It ended with a one-word comment ask instead of "link in bio." Every comment showed us which hook landed. What we changed after it: we now track saves first, write every script around one concrete comparison, and review the numbers every morning before writing the next one. What hook format has gotten you the most saves, as opposed to views?

A lead follow-up workflow can look healthy for days while nothing reaches anyone. The dashboard stays green, the API returns a message id, and the client finds out when a prospect complains. The three failure modes I see most: 1. "Accepted" is not "delivered" The send API answers with a success code or a message id. That only means the provider took the request. Delivery can still fail later: trial account limits, a dead webhook, a status code nobody is reading. If you only watch the send step, you miss weeks of undelivered follow-ups. 2. Silence looks like a quiet week A dead trigger or a broken webhook looks the same as a slow period. No errors fire. Volume just drops to zero. I treat "zero sends in a window where we normally send N" as a failure, same severity as a red error. 3. Timeouts that double-send, and skips you cannot see A timeout after the email already went out can make a retry fire a second copy. And a skip with no log looks broken to the client. Log the skips too. The daily check that catches this: one test message to my own number or inbox every morning, through the same path the client leads use. If that test does not land, I know before the client does. Second layer: alert when no status comes back within 10 minutes, and on failed or undelivered delivery callbacks. Thrown errors alone miss most of this. I also open-sourced the Claude Code skills I use for the rest of client work (scoping, quoting, outreach sequences). MIT, no email wall: https://github.com/CalebAnderson-Coder/ai-agency-os-starter Which silent failure have you actually been burned by on a live follow-up workflow?