On September 19, a company called Reinventing.AI put eight “AI employees” on GitHub under an MIT license — free, commercial use included, no signup. Eight roles: GTM engineer, SEO/AEO, web dev, social media, ad manager, sales, customer satisfaction, and a chief of staff. Between them, 59 routines.
Open one and there's no product. There's a folder of plain text files: a description of the role, an operating contract for what it's allowed to do, a schedule, and the routines themselves. An agent you already have — Claude Code, Codex, a handful of others — reads the folder and runs it on your own computer. It briefs you in the morning. Anything that sends, publishes, or spends money only happens on a channel you've explicitly released to it.
Being honest about the ceiling here: this is developer-shaped. If a terminal window makes you close the laptop, the download isn't for you this month. But the design is worth reading like a memo, because it quietly names the thing most business owners get wrong about AI.
You have a tool. You don't have a hire.
Roughly three-quarters of small business owners now use or are testing AI, and the ones paying for it report real returns — the average small business saves somewhere around five to six hours a week, owners a bit more. Good. But look at how almost everyone uses it: they open a chat window when they remember to. The AI is brilliant and completely passive. It waits.
That's why the time saved plateaus. You only get help on the tasks you happened to think of, on the days you happened to think of them. The work that quietly kills small businesses — the unchased quote, the unanswered review, the month you didn't post — is exactly the work you don't think of, because thinking of it was the missing piece in the first place.
What these eight folders do differently is put the AI on a calendar. Not “summarize this when I ask.” Friday at four, every Friday, find the reviews nobody answered. Same job, radically different result, because now it runs on weeks when you're underwater — which are the only weeks that matter.
An AI that helps when you remember to ask is a tool. An AI that runs Monday at eight whether or not you remember is closer to a hire.
Four things a real instruction has
Strip the repo down and every routine answers the same four questions. This is the useful export, and you can write it on an index card:
Who is this? One paragraph naming the role and what it cares about. “You handle follow-up for a two-truck plumbing company. You care about quotes that never got answered.” Not a personality — a job.
What is it allowed to do? The operating contract. Draft, don't send. Flag, don't refund. Post to this account, never that one. Most people skip this and then feel uneasy handing over anything real — the unease is the missing contract talking.
When does it run? The part everyone skips, and the part that does the work. A day and a time. If there isn't one, you don't have a routine, you have an intention.
What exactly does it do, step by step? Pull this list, compare it to that one, write the draft, put it here, tell me what you found. Specific enough that a competent stranger could follow it.
The version you can run this week
Skip the repo. Pick the one job that slips most in your business — for most owners it's follow-up, and it's not close. Write those four answers for it in a document. Then set a calendar reminder for one specific morning, and when it fires, paste the instruction into whatever assistant you already use.
Yes, a human is the scheduler in that version. That's fine. The scheduler was never the hard part — a dozen tools will automate the trigger once you know what should fire. The hard part is that almost nobody has ever written down what the job is, what it may touch, and when it happens. Do that once and you own something that survives every model release, every price change, and every tool you switch to next year.
Eight strangers' job descriptions went up for free this week. The reason they're worth looking at isn't that you should hire them. It's that they're proof the valuable thing was never the AI. It was somebody sitting down and writing the role.