How to be more ambitious with AI: the PM version
Nov 18, 2026
Zihan Wang
How to be more ambitious with AI: the PM version
Felix Rieseberg, EM at Anthropic, wrote the piece I wish I'd read a year ago: "How to be more ambitious: A tutorial for humans with models". His point is that most of us use AI far below what it can do, and the gap has less to do with the model than with us.
He splits ambition with AI into four kinds:
Speed: the same work, in less time.
Altitude: hand the model the outcome, not the task.
Depth: do the task the way the best expert in the field would.
Range: bring in fields the task doesn't obviously need.
Most of us live at Speed. I certainly did (hello, ticket writing agents). But his piece made me think hard about Altitude, and what AI can really do.
Altitude: hand over the outcome, not the task
Felix's idea is simple. Every task you give a model is your best guess at how to reach some outcome. So skip the guess and hand it the outcome. He used to ask Claude to fix individual crashes. Now he asks it to keep his apps crash-free.
This reminds me of the oldest advice in product: start with the problem, not the solution. We say it to stakeholders every week. "Don't tell me you need a dashboard. Tell me what user problem you are trying to solve."
Then we open a chat window and ask for the dashboard (sigh).
A task is a solution someone already picked. Handing AI the outcome, however, means going back to the problem, and letting the model question the solution with you.
Here's what that looks like in practice.
Instead of: "Write an interview guide."
Try: "We need to know if teams would pay for this by Friday. Plan the fastest way to find out, interviews or not."
Instead of: "Suggest onboarding improvements."
Try: "Get time-to-value under 60 seconds. That's doable. Work back from there."
Instead of: "Write the PRD for bulk export."
Try: "Trial-to-paid is stuck at 4%. Here's the funnel and 20 churn interviews. What should we build, and is bulk export even it?"
Instead of: "Write user stories for the sprint."
Try: "Get this shipped by the 15th with the team we have. What's the smallest version that still tests the bet?"
Instead of: "Write the launch plan."
Try: "Get 30% of active teams using it within a month. Plan backwards from that."
Make it a habit. You won't remember to do this every time. I don't. So save it once, in your custom instructions (or your CLAUDE.md if you use Claude Code), and let the AI ask for you:
Before you start any task I give you, check what it's for. Ask me what outcome I'm after, if I haven't said. If there's a more direct way to that outcome than the task I asked for, tell me before you start. Then tell me what data you'd need to do it well.
There's a catch. Every "try" above assumes the AI can see your funnel and your interviews. If it can't, you're back to pasting screenshots into a chat window. Getting your work where the AI can reach it is the unglamorous part, and honestly it's why we ended up building our own AI operating system.
Depth: work like the expert, not just talk like one
You may have noticed that "time-to-value under 60 seconds" sounds a little unreasonable. That's on purpose: set a goal that sounds unreasonable, and tell the model it's reasonable. Demanding leaders have done this to their teams forever. Now we can do it to our models.
His bigger point on Depth: ask a model to "act like an expert" and it tends to change its vocabulary, not its method. We're literally getting jargon-dumped. So ask for the expert's actual tools and the raw data, and never let the model grade its own homework.
If you're not sure what "deep" looks like, Felix shares a prompt for exactly this in his piece, and it's gold. Here's our version of it:
Before you start, help me go deeper on this task without changing its scope. Answer these, specific to this task:
1. What would the best professional in this field do here that I'm not asking for?
2. What raw material would they start from, instead of something already summarised?
3. Which methods and tools would they use?
4. What, other than you or me, would show the result is right?
Then rewrite my prompt to ask for exactly that.
Want it as a one-word command? It's free as Go Deeper in our AIOS template. Credit for the idea goes to Felix.
If you found this useful, please share it with your friends.
Back to Overview