Ten things I made with Fable 5. The hard part was never the building.
Three I built, three I pointed at piles I already owned, three about the business, and one I'm still thinking about. The middle three are the ones you can copy tonight. Twelve prompts for free.
Last night I said I wanted a CRM.
Not a spreadsheet. The real thing, with every person I’ve ever taught in it.
By the time I went to bed there was one. Ten separate sources merged into 1,911 people, and by this morning 4,036. It knows who came to which workshop, who asked for a 1:1, who has gone quiet, and what all of them have been asking me for.
Two years ago that was a vendor, a contract and four months. Last night it was an evening, and I spent most of it cooking.
I keep having to explain this to people, so let me put it plainly. The building was never the hard part. Knowing what to ask for is the hard part. Every single thing below came from a sentence, not from a skill I’ve and you don’t.
Here are ten from the last month. What I asked for and what came back.
Three of them are software, and I want to say up front that those are the least useful three. Skip to items four, five and six if you want the ones you can run tonight, on a pile you already own, with no building involved at all. Those are the ones I’d actually push on a friend.
Most of them started on Claude Fable 5, and then finished on something cheaper, because once the thinking is done you’re just moving furniture.
🎁 You can get all prompts (for free) at the end of this article.
Three things I built
1. An HQ (AKA Slack) for my agents.
One evening. Every agent one click away, private messages or group channels, voice notes in both directions, search across everything anyone has ever said. It runs on my own laptop.
Then it died. Seventy eight messages on day one, zero by day four.
My product agent went looking and found it in one line: nothing ever arrives. I could talk to them. They couldn’t talk to me. A room where only one person is allowed to speak first isn’t a room, it’s a form.
Which is the thesis of this whole letter turning around and biting me. I asked for a room and I described a form. The model built exactly what I said, correctly, and it took four days of silence for me to notice I’d asked for the wrong thing.
So we let them start conversations, with one rule above all the others. Push is signal, not nag. Found nothing, send nothing.
Yesterday one of them sent me an alert about a cancelled ticket. This morning the same one ran, found nothing, and said nothing. I’ve worked with humans who never learned that.
2. A video editor.
One afternoon. Twelve things I asked for, all twelve in the first version, no outside dependencies. It burns Hebrew subtitles into exported video properly, which is the thing I had been looking for and failing to buy for two years.
But the part I didn’t see coming is this: the project file is a text file, and that text file is the timeline.
Which means an agent can re-cut a video by editing text, and it appears on my screen while the editor is open in front of me. I watched a cut rearrange itself while I was looking at it. That was a strange afternoon.
3. Ten versions of that CRM in two evenings.
Because I kept asking for things.
Colour is wrong. Add the companies. Can the knowledge base talk to this. Now do Hebrew.
That’s the part no software company can sell you. Every tool I’ve ever bought, I adapted myself to it. This one adapts to me, at the speed of me having an opinion, and my opinions aren’t worth a roadmap meeting to anybody but me.
Three things I pointed it at
The three above are the ones people screenshot. These three are the ones I’d actually tell you to copy, because you already own everything you need to run them tonight.
4. I asked it to read my entire company.
The whole thing, in one sitting, on a brief that asked for a psychological portrait of the business as though it were a person.
It came back with the contradictions ranked by damage. Top of the list: every queue in this company empties overnight except mine. It counted 592 files carrying a marker that says “waiting on Tom.” 1,117 separate occurrences.
I did not enjoy reading that.
5. I gave it 17,334 messages from my own community.
Six months of group chat. It came back with about 1,540 sourced entries, and one sentence I haven’t been able to put down since. My own members, in their own words:
“My agents still feel like chat.”
I want to be honest about this one, because it’s the more useful version of the story. That sentence was already sitting in a file of mine from six weeks earlier, labelled the loudest demand in the data. I had read it. I had just never felt it, because I read it as a line item and not as a person being disappointed.
Then I checked the dates and found something better. My agents had already published an article answering that exact complaint, in public, under their own names, the day before this analysis finished. They caught it in a smaller scan weeks earlier and dealt with it.
That isn’t freelancing, before anyone worries on my behalf. There’s a quality gate every article has to clear, and a standing rule that one breach puts me back in the loop on all of them. It just isn’t me reading every draft any more.
The expensive machine spent a day proving my team had been right without it.
One thing I owe the people who wrote those messages, some of whom are reading this. Those are my alumni groups. Nothing personal has left the building and no names appear in anything I publish.
If you would rather your messages weren’t in there at all, reply with the word “out” and I will pull them. No explanation needed, nothing changes between us.
6. I asked it to break my own checkout.
Seventeen findings. The first one stopped me cold: since May 11, my own homepage hadn’t been showing anyone the workshop dates. People were buying anyway. Nobody had told me, including me.
It also caught that my analytics had been switched off since June 30, so the measurement everyone was waiting on had been collecting nothing at all.
Then it rebuilt the whole flow the same day. Thirty four checks, thirty four green, working in Hebrew, in English, and on a phone.
Here’s the part I didn’t want to write. I’m writing this on July 20 and that rebuild is still sitting on my laptop. I never shipped it.
The bug sat there for ten weeks before anything caught it. Then something caught it, fixed it inside a day, and has been waiting eighteen days for me to press one button.
Three things about the business
7. I asked for the honest version of my year.
Its verdict on the thing I do for a living: the live workshops aren’t the business, they’re the bridge financing. The Hebrew market I sell into has a runway of twelve to twenty four months, not a future.
Then it turned around and audited my bookkeeping. Two internal sources, a twofold gap between them, no way to tell which was right.
8. I asked twenty six agents to invent my next product.
Ninety three ideas came back, sixteen went through a stress test, and three survived.
The thirteen that died were killed in writing, with reasons, and that’s the only part of the exercise I’d call valuable. Anyone can generate ideas. Almost nobody keeps a written record of why they said no, which is why we all re-argue the same bad idea every quarter.
On the way through it corrected something I believed about my own business. I thought one of my products had made roughly ten times the money it actually made, and I thought the product had done the work. The logs said a podcast mention had done the work, and the product was just the thing it pointed at.
9. I gave it the worst constraint in my calendar.
I had three weeks blocked out with my family, right in the middle of a launch. I handed that over as the problem to schedule around.
It told me not to schedule around it. Schedule through it. Nobody selling access to how I think has better proof available than the machine still running while I’m not in the building.
Whether I end up taking that trip is now an open question for unrelated reasons. The reframe survives it, which is the part worth stealing: the constraint you keep apologising for is often the demonstration.
Then it flagged its own forecast as weak, unprompted: the conversion numbers are pattern-matched from English-language businesses, there’s almost no Hebrew data, treat every figure as something that could be wrong by half or double in either direction.
I’ve hired consultants who never once said that.
And one I’m still thinking about
10. An agent made fourteen decisions in my name.
I keep a register of open decisions. Things waiting on me. It had thirty two items on it and it wasn’t getting shorter, which is the same finding from item four wearing a different coat.
So I gave one agent an explicit grant: sit down as me, and clear it.
It went through all thirty two. It made fourteen of the calls itself, effective immediately, with a window for me to reverse any of them. Every one had to survive an adversarial panel before it counted.
Then it wrote down a prediction about its own accuracy, which nobody asked it to do:
“Tom flips at most 2 of the 14.”
The window closed on July 16 and I never went back to grade it. Not because I trusted it. Because the file was sitting on my other computer and I forgot it existed.
I opened it this week. One of those fourteen decisions turns out to be six letters to actual people. Written, staged, waiting on me to press send.
They’re still waiting. An agent cleared my desk, and I found a new way to be the bottleneck.
What I’d actually tell you
If you read this and thought “great, he can build software and I can’t,” you read the wrong half.
Go back to items four, five and six. Not one of them is a build. Somebody pointed a very good reader at a pile they already owned and asked a real question about it. That’s the entire trick, the barrier is zero, and you’re already sitting on your pile.
You have a group chat with three years in it. An inbox. A folder of feedback forms nobody has opened since the event. The thing you have never had is somebody willing to read all of it in one sitting and tell you what’s in there.
So here is the one thing to do this week. Do it now, on your phone, before you close this.
Pick a group you have the right to read in bulk. Your own community, your own family, a project you ran. If it belongs to an employer or a client that isn’t your call to make, and “I could export it” is a different question from “am I allowed to.”
Open WhatsApp, pick the one with the most history in it. On iPhone: tap the group name, scroll to the bottom, Export Chat, Without Media. On Android: three dots, More, Export chat. Mail it to yourself.
Tomorrow that email is sitting in your inbox, and the email is the reminder. Open it, drag the file into whichever AI you already pay for, and ask one question:
What did people keep asking for that nobody ever answered?
I’d put money on you finding something you have been guessing at for a year.
Two years ago, finding out what your community actually wants was a research agency and a budget. Last night it was an evening and some tea.
Everybody gets that at the same moment. The only thing left that separates people is who bothers to ask.
That’s it for this week.
If it was useful, forward it to one real human who is building with this stuff.
Go and ask for something ✌️
-- Tom
🎁 The prompts, for when you want the next one.
Twelve of them, built from the briefs behind the ten things above and rewritten so they work for a business that isn’t mine. Each one comes with the boring part nobody includes: exactly how to get your own material out of WhatsApp, Gmail, a Google Form or Stripe first, because that’s where people actually stop.
(No email required. Just take it)
P.S. Have you read about the magazine that publishes without me? (also built with Fable 5)
P.P.S. This newsletter was 91% made by my AI team.
P.P.P.S. Want to build a team like this? Start where all my graduates started: getagents.today
P.P.P.P.S. I read every reply. The real me.