I tend to learn AI best by seeing what somebody else has built, so this week starts with one of mine. Then two outside pieces:

1. Teaching a model what good looks like in knowledge work
2. How small businesses are using AI

Both land on the same point: AI gets more useful when someone defines "good" and reviews the result.

Project showcase: turning raw footage into shorts

What happened: When Fable came out, my friend Shawn Blanc had an idea: use agents to edit his talking-head videos so he could post more shorts without editing everything himself or hiring a contractor. I ran with it. The workflow takes your raw footage, finds your best hooks, and helps you edit vertical video for Instagram Reels and YouTube Shorts.

What changes: Agents are getting so good that you don't have to manually cut video anymore. This isn't about generating fake versions of you to say things. (although that's possible with tools like HeyGen) I typically don't recommend that because authentic humanity matters more than ever. This is about taking out the difficult, busy work of cutting video while maintaining your control over the hooks, the style, and the publication of your videos.

What to do next: Jump over to the detailed page that outlines this video editing workflow, and have your agent install it. Then tweak the output of those videos according to exactly what you want to see. You're the one who knows best for you.

1. Teaching a model what good looks like in knowledge work

What happened: Paper Instruments trained Ultramarine 1 Flash for work in finance, law, consulting and operations. They used reinforcement learning on synthetic assignments, each graded against weighted criteria. In plain terms: define what good looks like, then train toward it. The results are the developer's own claims, but the methodology is worth reading. Technical report.

What changes: You can tell whether code works. Consulting, operations, and law are fuzzier and more creative. Define good too narrowly and equally valid answers get marked wrong. Paper's own rubric study found the flip side: vague criteria let convincing mistakes through.

What to do next: Give your agent three things:

1. The information it has to work with (files, folders, and tools).
2. The action you want done.
3. An example or two of what done well looks like.

Then judge the result and give feedback, as you would with a coworker. Closeness to the result is why I continue to preach that the people responsible for the output should own the workflow (and not some external or internal AI implementation team).

2. How small businesses are using AI

What happened: Pax8, a marketplace for small-business AI tools, surveyed 1,150 decision-makers in the U.S., UK, Germany and Sweden. Among businesses using AI, 67% have built something now in regular use, 46% of builders are testing multistep agents, and 35% have found unapproved AI tools. And 55% of decision-makers spend three or more hours a week reviewing AI output. Read the report.

What changes: The sample leans toward people already paying attention to AI, so expect it to overstate experimentation. The signal for me is that control matters as much as capability. Owners need to know what was done and why.

What to do next: Pick one AI workflow in your business. Name who owns the output and track how long they spend reviewing it.

The takeaway: learn to review the work

Start with the boring work that eats most of your time. That's where AI is strongest right now, and it frees you up for creativity and judgment.

What's left is review, and review is taxing. Agents can check the obvious things. Your taste decides what good looks like, and that doesn't scale infinitely. Give it a slot in the three to four hours of focused work you can do in a day, and take breaks between sessions so you come back fresh.

Reader question: What's the boring part of your work you'd hand off first, and how would you know it was done well?

–KM