Honest automation
On getting the leverage of a scheduled AI job without giving up the say over what it does.
Say you let a scheduled assistant handle something that goes out every week. It takes the repetitive part off your hands — genuinely useful. What you don’t want to hand over along with it is knowing what got said, and getting to approve it first. Splitting those two — automate the task, keep the decision — is the idea CRAFT for Claude Cowork is designed around.
The moment you hand a job to a schedule
There’s a real pull to it. You have something that repeats — a weekly summary to post, a report to pull, a message that goes out every Monday — and an assistant that could, in principle, just handle it. The appeal is obvious: you stop doing the boring part. The hesitation is just as reasonable. Once it runs on its own, on a day you’re not watching, what exactly did it say in your name? You find out when everyone else does. That trade — less effort, less certainty — is the one most “AI does it for you” setups quietly ask you to make.
The version that should make you nervous
The setup that deserves the suspicion is the common one: a model composes something in the moment and ships it, unattended, and the first you hear of it is when it’s already out. Reach across tools and the feature line reads the same — “scheduled tasks,” “automations,” “let AI run it for you” — and side by side the checklists look identical, the way two cars both list cruise control. What the line doesn’t tell you is the part that matters at two in the morning when the job fires: who decided what it would say, and how you’d know if it went wrong.
The difference is the shape
CRAFT’s version isn’t a promise that the automation is smarter or won’t slip. It’s an arrangement where the nervous-making part isn’t on the table. You approve the exact thing before it can run — the specific text, the specific action — and approving it freezes it. Each cycle delivers the item you approved for that cycle, so freezing pins the exact words without pinning you to one message forever; next week’s post is next week’s approval, whether you queue a few ahead or approve each as it comes. The scheduled job composes nothing on its own; its whole permission is to deliver the frozen, approved item when it’s due, and otherwise touch nothing. And every run leaves a receipt — including the runs where nothing was due — so a job that quietly stops firing shows up as a gap in the record instead of a silence you never notice. The record doesn’t read itself, of course; what it does is turn a silent failure into something you can catch the next time you look, instead of something you can’t. “Automated,” arranged this way, stops meaning “unwatched.” You’re still the one who decided; you just didn’t have to be awake for the delivery.
Two ways to hand a job to a schedule: compose-and-ship on its own, or deliver only what you already approved — and leave a record either way.
A run receipt — structure real, details invented
2026-08-31 09:00 run: weekly-post status: POSTED ← item approved#142
2026-09-01 09:00 run: weekly-post status: NOTHING-DUE
2026-09-02 09:00 run: weekly-post status: POSTED ← item approved#143
The format is illustrative; the habit — a line for every run, including the empty ones — is the real one.
Why it’s built this way
This reflex comes from somewhere. CRAFT is built by one person who brings a safety-first mindset from a day job in cybersecurity. In that kind of work you learn not to trust an automated job you can’t audit — not because the job is untrustworthy, but because a process you can’t check is one you’re taking on faith, and faith is not a control. So you instrument it: you make every run announce itself, and you treat a missing announcement as a finding, not a non-event. A scheduled task that leaves a receipt for every run is that habit, written into a framework. Provenance, here, is accountability — the discipline exists because someone who does this for a living chose it, not because the automation is beyond going wrong.
A process you can’t check is one you’re taking on faith — and faith is not a control.
We run it on ourselves, in public
The honest way to stand behind “trust-but-verify” is to be caught doing it. Some of the posts on our own social accounts are published by exactly this kind of scheduled job — approved in advance, frozen, delivered from a queue, with a receipt for every run — and we document how that goes in the open, including the weeks it misfires. A companion post is us eating our own cooking. This one is about what the same shape gives you: a way to let a schedule carry the work without carrying off the decision.
Put it around your own work
CRAFT for Claude Cowork is free during beta and lives on GitHub — no login, no email, no paywall. It runs on Claude Cowork’s scheduled tasks, and it brings that same discipline to the automated work you build on it: a job can run while you’re not looking, and you can still see exactly what it did afterward.
The quickest way to test any of this: point CRAFT at one job that repeats, let a run happen while you’re away, then read exactly what it did. It’s free to download during beta, and approving-then-verifying is simply how it handles work you schedule.
Free during beta. The download is ungated — no account, no email, no checkout. CRAFT for Claude Cowork runs on Claude Cowork, so you’ll need Cowork available in your Claude plan; CRAFT adds no cost, account, or gate on top of it. More on what it is
CRAFT for Claude Cowork is an early framework, in beta, and AI can make mistakes — review anything that matters before you rely on it. Automating a task this way keeps you the one who approved it and leaves you a record to check; it doesn’t make the automation incapable of failing, and the receipts exist precisely because it can.
