User
Write something
Pinned
Start here. What this group is, and who it is for.
This group is a build log. Every week I build one automation that does a real job, and I publish all of it: the brief I wrote before starting, the prompts, the files it produced, the review step that caught the mistakes, and an honest note on what broke and how long it took. I publish the failures because they are the only part that tells you anything. Anyone can show you a working demo. What you cannot tell from a demo is what happens when the job reports finished and is not, which is the thing that actually decides whether this stuff is worth paying for. Who I am writing for. Owners and operators who are paying people to move information between systems. Invoices retyped into an accounting package, intake forms into a CRM, claims into a portal, two systems after a merger that have to become one. If that is your cost base, this log is about your problem. If you want to build these yourself, everything here is open and you are welcome. I am just not writing it for you. What to do first. Read the most recent build, then open Classroom and read "When It Breaks". That one is the honest centre of the whole thing: it is the three ways an automation lies about having worked, and how each gets caught. If you only read one page here, read that one. What I will not do here. I will not sell you anything in this feed, I publish no income figures, mine or anyone else's, and I do not do "AI strategy". I build one thing at a time and show the receipts. Comment with the process in your business that you would most like to never do by hand again. I read every one, and the ones that come up repeatedly become the weekly build. Make AI do the work.
0
0
The check I now put in every job: make it refuse
Most automations do not fail loudly. They fail quietly and keep reporting success, and you find out weeks later from a customer. The cheapest guard I know is a check that refuses. Pick the one number the job cannot sensibly produce. Usually it is zero. A job that files invoices should never finish having filed zero invoices on a Tuesday. A job that pulls yesterday's orders should never return an empty list on a trading day. A job that reads a folder should never find nothing in a folder that always has something. Then make that case fail on purpose. Not a warning in a log nobody opens. Fail, stop, and tell somebody. The reason this one earns its place: the silent failures are the expensive ones, and almost all of them look exactly like a normal successful run with nothing in it. An empty result is the most common shape of a broken job, and it is the one that looks most like a quiet week. One line to write, and it turns the worst failure mode into an email you get the same morning.
0
0
I thought I had locked a page. A stranger could read the whole thing.
A failure from my own week, published here because the failures are the only part that tells you anything. I set two courses in this classroom to members-only. The setting took. The interface showed the lock. I had the confirmation on screen. Then I loaded the page the way a stranger would, signed out, in a clean browser, and read the entire thing top to bottom. The lock was decoration. The group is public, and public beats the per-item setting. The confirmation I trusted was describing a field in a database, not what a human being could actually see. I took both pages straight back down. The general version of this is the thing worth keeping. A setting that reports success is telling you it was saved, not that it works. The only check that counts is the one made from where the customer stands, in their browser, with their permissions, not yours. Same for an automation that reports finished: finished is a status it wrote about itself. So the habit, for anything you build or buy: before you believe it, go and look from the outside. It takes two minutes and it is the difference between a thing that works and a thing that says it does.
0
0
What one reporting job actually cost, counted properly
Somebody asked me to automate their reporting. Before writing anything I counted what the job was, which took about fifteen minutes and changed the answer. The numbers: 23 live ad accounts, in 4 currencies. Another 13 accounts that were still on the list and had been dead for months, which nobody had noticed because nobody had ever counted them either. One minute per account per cycle is 23 minutes a cycle. Run that every week and it is close to 20 hours a year, on a job that produces no decision most weeks. Here is the part that surprised the owner. The expensive part was not the reporting. It was the 13 dead accounts quietly making every report longer, every month, for a year. Counting first is free and it is boring and almost nobody does it. If you do nothing else this week, pick the job in your business that happens most often and multiply: how many times a cycle, how many minutes each, how many cycles a year. If the number is small, you have saved yourself a project. That is a good outcome and it is most of them. The full costing is in the classroom under What this actually does. No sign up for it, it is just there.
0
0
1-4 of 4
powered by
Make AI Do The Work
skool.com/ai-automations-3776
AI automations and AI agents that actually run. One working build published every week: the brief, the prompts, and what broke.
Build your own community
Bring people together around your passion and get paid.
Powered by