Stop prompting, start managing
Ethan Mollick's guide to picking an AI now that they do whole tasks: think about the model, the app and the harness.
18 Feb 2026 · 4 min read
By Ethan Mollick (One Useful Thing) on Newsletter
The short version
- Choosing an AI now means choosing three things: the model, the app and the harness.
- Pay for one good plan (around $20 a month) and pick the strongest model yourself, not “auto”.
- Try it on real work you need doing, not test questions.
- Using AI now looks more like managing a person than writing clever prompts.
Who it's for: Anyone who feels behind on AI and wants one clear place to start.
Why it's genius
Ethan Mollick is a Wharton professor whose One Useful Thing newsletter is one of the most-read guides to AI at work. His guides to “which AI to use” come out every so often, and this one changes the whole way of choosing.
He splits it into three parts. The model is the brain. The app is where you use it, like a website, phone app or desktop tool. The harness is the set of tools that lets the AI take actions and finish multi-step tasks. The same brain can do very different work depending on its harness.
His bigger point is that using AI now looks like management. You aren't writing clever prompts any more. You're handing over real tasks, then watching and steering.
Try it
Pay for one of ChatGPT, Claude or Gemini, and pick the strongest “thinking” model yourself rather than “auto”.
Give it a real piece of work you need this week, not a test question.
Push back on its first answer, like you would with a new colleague.
Try NotebookLM with your own documents to see a different kind of harness.
Here's a task I'd give a new colleague. Ask me any questions you need first, then tell me your plan before you start.
Our take
Why the harness matters more than people think
Most “which AI is best” arguments are about models. Mollick's point is that the harness can matter just as much. A strong model in a plain chat box can only talk. The same model in a harness that can browse, run code or edit files can actually do the job.
That's why the same AI can feel brilliant in one app and useless in another. When something disappoints you, it's worth asking whether it's the brain or the tools it was given.
The free version is holding you back
His most practical advice is blunt: pay for one plan and choose the model yourself. Free versions are tuned for speed and chat, so they make more mistakes. The default “auto” choice often picks a weaker model too. For hard work, pick the thinking option.
He's also honest about the risks. Some tools are research previews that eat through usage limits, coding tools carry risk for people who don't code, and he warns firmly against giving any AI broad access to your computer and accounts.
Managing is the skill to build
If AI work is management, the skills are the ones good managers have: a clear goal, the right context, sensible chunks of work and honest feedback. None of that needs technical knowledge, which is good news for people who feel left behind.
It's the same theme as Anthropic's meta-work research, from the other side. Anthropic measured how much time managing AI takes. Mollick tells you how to get good at it.
Where it shines
People who feel behind and want one clear, honest place to start.
Where it breaks
Model names go out of date fast. The framework lasts, but the specific picks won't.
Where Ugenie helps
Once you're managing several AIs, you need someone keeping notes. That's Ugenie, and the AI Council asks them all at once.
Get Ugenie →Tools mentioned
ChatGPT, Claude and Gemini
The three leading general AI assistants, from OpenAI, Anthropic and Google.
- Cost
- Free versions exist; the useful paid plans start at about $20 a month.
- Our verdict
- All three are excellent and they leapfrog each other every few months. Pick one, pay for it, and learn it well. Ugenie's AI Council asks all three when it matters.
NotebookLM
Google's research notebook. Upload your own documents and ask questions about them, with answers tied to your sources.
- Cost
- Free, with paid tiers for heavier use.
- Our verdict
- The easiest way to see why the harness matters. Great for reports, contracts and research you need to actually understand.
Claude Cowork
Anthropic's desktop agent for everyday work like files, documents and research, rather than code.
- Cost
- Comes with paid Claude plans.
- Our verdict
- Promising for people who don't code. Mollick notes it was a research preview at the time and can use up your limits quickly.
Prices are rough and change often. Check before you buy.
Questions people ask
Which AI should I use?
Pick one of ChatGPT, Claude or Gemini, pay for a plan, and choose its strongest model yourself. They're all very good; learning one well matters more than picking the “best” one.
What is an AI harness?
The tools and setup around an AI model that let it take actions, like browsing, running code or editing files. The same model can do very different work depending on its harness.
Is the free version of ChatGPT good enough?
For quick questions, yes. For real work, Mollick recommends a paid plan, because free models are tuned for speed and make more mistakes.
All credit to Ethan Mollick. We share things we think are genius and add our own take. Watch or read the original for the full thing.