Four AI sessions worked my night shift. The work was done when I woke up

Happy Wednesday.

While I was asleep, four AI sessions kept working on my company's marketing. This issue is about what made that possible, which is nothing fancier than written instructions inside Claude Cowork. What these new AI work platforms are, why they sit unused inside most businesses that buy them, and how to write the first instructions page for Cowork yourself this week. One page. No code.

Also in here: three things worth your time and one AI word decoded.

At 10:30 last night I wrote four one-page briefs and handed each one to an AI session on my computer. Then I went to bed.

By the time I woke up this morning, the work was done. One session had drafted an outline of this week's LinkedIn posts, every fact pulled from a source file, images built. One had turned my nine-day marketing plan into a planning board I can run daily with minimal effort. One had cut a short video ad; I'd seen the first version before bed and said no, and a rebuilt cut was waiting with the others. It still wasn't perfect, but I was 90% there. The fourth had upgraded one of my internal review tools. Before I went to bed I recorded my screen for a minute: four panels, all quietly working.

Every business can do this right now for $20/month and desktop software available for free. Anthropic, the company behind Claude, sells it as Claude Cowork. That's the one I mostly run. OpenAI sells its version as ChatGPT Work. The difference from the chatbot you've already tried: a chatbot answers your question and leaves the doing to you. These platforms do the doing. Hand one a job and a folder of files, and it reads, drafts, checks, and has the work waiting when you come back.

So you'd expect every business to be running one by now. Almost none are. Daf Dent, an industry writer, wrote about that gap in Marketecture this week: his own headline calls the platforms revolutionary and asks why nobody is using them. Twice he gets to the instruction files the platforms run on and asks, "But whose job is it to create and distribute them?" Twice he moves on without answering.

I can answer this, because it's most of what I do all day. The job is yours. Not the typing, the deciding. A platform stalls on decisions your business has already made but never written down. Which tools may open which accounts, and who decided. Where the files live, and who may read them. How each recurring job is done, and what counts as done. Who may change any of that once it works. Those decisions are the operating instructions of your business, and no vendor can ship them in the box, because they're about your company, not their product.

Make it concrete. Say you want a platform to handle your customer follow-up. Before it can, somebody has to decide which inbox it may read, whether it may see the billing system, what a finished follow-up looks like, which customers are too sensitive to automate, and who signs off before anything sends. Five decisions. Every one is yours, none is technical, and the platform cannot start until they're written down. Multiply that by every job you'd want handled, and you can see where the stall comes from.

The unwritten version already exists. It lives in your head, in your office manager's habits, in the way the person who's been there nine years knows which invoices get held until Thursday. That works, until you hand the job to someone who hasn't been there nine years. A new hire gets weeks of shadowing to absorb it. The software gets a login.

My four overnight briefs were those decisions, written down: what the job is, what done looks like, what the session may not touch, where the result goes. On Monday I counted twenty-seven AI sessions running on my computer across the day, and every one was following a page like that. Claude Cowork and ChatGPT Work come without instructions. The businesses writing them down and combining them into a system are the ones pulling ahead.

Two kinds of pages made last night work, and your business already runs on both. The four briefs I wrote before bed were work orders: do this job tonight, here's what done looks like, leave everything else alone. The same thing you'd write when handing a job to a contractor. The other kind is the standing page: how a recurring job gets done, written once, followed every time the job runs. My company runs on more than sixty standing pages. One drafts LinkedIn post outlines. One turns a raw meeting transcript into a client recap with action items. One keeps my task list honest. Each is a page or two of plain language, and not one of them took a programmer to write.

The two work together. A work order doesn't restate how a job gets done. It points at the standing pages that already say so. The brief behind last night's LinkedIn post named the job and what done looked like. The voice rules and the sourcing rules it ran on were standing pages, already on file.

Here is that standing LinkedIn page, abridged to its five headings, so you can see there's no magic in it:

Notice how much of that page is decisions and how little is technology. Which sources count. What gets flagged. Who signs off. I made those calls once, and now every run follows them, at ten in the morning or ten at night.

The "who checks it" line earned its keep this week. That video ad from last night: the first cut was weak, and I said so just before midnight. The brief already said nothing ships without my sign-off, so nothing did. The session took the rejection and rebuilt the ad while I slept. A written gate works even when you're not watching it.

Four rules keep sixty of these pages trustworthy, and none of the four requires a programmer.

1. Written once, in one place. When a copy goes out to where the work happens, the copy is never edited. Fix the original, reissue the copy. Otherwise you end up with six slightly different versions of a job and no idea which one ran.

2. A list says what's managed. One page names every part of the business the system may work in. Anything not on the list, it leaves alone. That list is also the answer to the question every owner asks first, "what can this thing touch?" You wrote the answer, so you can read it.

3. Every copy gets a receipt. When an instruction goes out to do its work, the copy is logged with a fingerprint, so a stale copy shows up as stale instead of quietly running last month's version of the job.

4. Written rules say who may change what. They bind the AI the same way they bind people. The line I repeat most often to clients: if nobody has access to a folder, the AI doesn't have access to it either.

None of this is code. This is management. Deciding who does what, what good looks like, and who may change the process is work you already know how to do. You have written instructions for closing the register and running payroll. This is the same discipline, pointed at the work that crosses your desk.

So here's this week's move: create one page. Pick one recurring job you wish you didn't personally own. Write the five headings from my LinkedIn page: what starts the job, the steps in order, what done looks like, who checks it, who may change this page. Fill them in. I've made it a printable sheet, the five headings ready to fill in, my LinkedIn page on the back as the worked example: download it here. That page is useful the day you write it, with no AI anywhere near it, because you can hand it to a person and they can run the job. And the day you do turn on a work platform, that page becomes the first instruction it follows. The platforms call a standing page like this a Skill. The last section takes that word apart.

Now the part of this issue I've been looking forward to. This is the material I teach in private team trainings, and to the teams inside the companies where we install full AI systems. Starting next Thursday it runs as an open cohort for the first time: ten business owners, live, each building on one real workflow from their own company. You pick the job, you write its pages with my help in the room, and you leave with the workflow running.

And a scheduling note worth saying plainly: next Wednesday's issue lands the day before the first run starts, so if Thursday is your run, this is the issue to act on. Reserve a seat, or read what the four weeks cover first.

Small businesses adopted AI faster than they wrote rules for it. Intuit's 2026 AI Impact Report drew on more than 34,000 small and mid-sized businesses; 77% use AI daily. Asked about a written AI policy, owners answer with what one Forbes contributor calls a blank stare. His diagnosis: the tools arrived one subscription at a time, and nobody wrote the rulebook. A separate study he cites found 49% of employees using AI tools their company never approved. The rulebook everyone skipped is the page this issue is about.

More small business owners now ask an AI about their insurance than ask an agent. In a survey of 501 US owners published August 18, 36% had used an AI chatbot or search tool to explore business insurance, against 31% who consulted an agent. Of the owners who used AI for insurance advice, 94% said they trust what it told them. The survey is an insurer's own, so read the framing with that in mind. The number worth sitting with is the 94%. Trust at that level with nothing underneath it is why "who checks it" is one of the five headings above.

A court has started sorting out who pays when an AI agent goes somewhere it should not have. A US appeals court ruled on August 5 that Amazon was unlikely to win its claim that Perplexity's shopping agents broke the federal computer-fraud law by acting inside customers' own accounts. The Reuters explainer names the part that reaches ordinary companies. The business that deployed an agent can be sued, not only the business that built it. None of it is settled, and it does not need to be for the lesson to land. Write down what your AI is allowed to touch, and who signs off before it acts.

The first word you'll meet when you open one of these platforms is Skills. A Skill is a saved instruction file: the standing page from this issue, stored where the AI reads it before it starts the job. Two things the word hides. The platforms like Cowork and ChatGPT Work ship with empty slots, and the vendor is selling you the filing cabinet; the files are yours to write. And whoever writes them decides how your business does that job, every time it runs. That's a real responsibility hiding behind a word that sounds like a menu feature. When a vendor demos Skills, the question that sorts a real offer from a canned one: "who writes ours, and who may change them?"

Write the one page this week. Then reply and tell me which job you picked. I read every reply.

See you next Wednesday.

P.S. I also teach this. Four weeks, live, ninety minutes a week, where you bring one real workflow of your own and leave with it running. Ten seats a run. The first starts Thursday 27 August, and there's a second on Wednesday 2 September if that week suits you better. $1,295 either way. Reserve a seat, or read what the four weeks cover first.