AI Rep-Loop Compression
Build AI tools that give you feedback 80% as good as an expert's, on demand, to get years of judgment-building reps in a fraction of the time.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 4
- Confidence
- 85%
A method for accelerating skill and judgment development using AI. Traditional learning (e.g. an analyst grinding for two years) builds judgment through slow, inefficient loops: do the work, wait for feedback, maybe get good feedback, try again. AI lets you compress both the speed and the number of those loops. You build custom GPTs that emulate an expert's feedback, or that generate personalized practice scenarios, so you can get high-quality reps on demand — infinite times — instead of waiting for rare real-world occasions.
Origin
Developed by Hilary Gridley, who detailed it on Claire Vo's 'How I AI' podcast. Prompted by the debate over whether AI eliminating entry-level grunt work will starve future professionals of judgment-building reps.
Core principles
- 01Entry-level grunt work is inefficient by design; its real value is the reps that build judgment and taste, not the output.
- 02The traditional feedback loop (grind, wait, get feedback, retry) is slow and inefficient — AI can shrink both the speed and the number of loops.
- 03A GPT that 'thinks like' an expert can deliver feedback 80% as good as theirs, on demand, without waiting for a one-on-one.
- 04Personalized, scenario-based practice makes learning more fun and can be tailored to your exact domain.
- 05We still must invest in people's skills — just not the historically inefficient way.
How to run it
- 1
Identify a skill limited by slow feedback
Find a capability you build only through rare reps — giving a t-shirt-size engineering estimate, making a logical argument, getting a manager's feedback — where opportunities come up only a couple of times a week or less.
- 2
Encode the expert's judgment into a GPT
Build a custom GPT with a prompt that captures how the expert (e.g. you, as a manager) thinks, so reports can get feedback close to yours whenever they want.
Pro tip Feed in your accumulated notes on how key people think and ask the model to infer the criteria they'd use — LLMs are strong at this pattern-matching.
- 3
Generate personalized practice scenarios
Have the GPT produce domain-specific practice — e.g. LSAT-style logical-reasoning questions set in realistic PM scenarios, with multiple-choice answers and explanations of why each is right or wrong.
Pro tip Make scenarios hyper-specific to the learner's actual company and domain (e.g. consumer health at Whoop) to sharpen relevance.
- 4
Run the loop at volume
Practice on demand, as many times as you want — spend an afternoon getting the reps a job would take months to surface naturally.
Pro tip The gain is both faster loops and far more loops than real work provides.
In the wild
To build the fundamental skill of logical reasoning that underpins strong argumentation, Gridley built a GPT ('Aristotle') that creates LSAT-style questions set in scenarios a PM would actually encounter — e.g. sales wants feature A, engineering only has time for feature B, and the metrics show a retention lift. It presents the logical relationships as multiple choice, you pick one, and it explains why you're right or wrong.
→ A product builder can get the kind of judgment-building reps that normally arise only occasionally on the job, unlimited times, in a fraction of the time.
Common mistakes
Assuming reps must take years
The belief that judgment requires a two-year grind ignores that the underlying loop — do, wait, get feedback, retry — is itself inefficient and can be radically compressed with AI.
Confusing the output with the learning
Fearing AI eliminates entry-level work misses that the value was never the grunt-work output but the reps; those reps can be delivered better and faster by AI.
Is it for you?
Best for
Individual contributors and managers who want to accelerate their own or their team's judgment and skill development, especially in fields where real-world reps are slow to accumulate.
Not ideal for
Skills requiring genuine high-stakes real-world consequences or human relationship dynamics that a simulated loop can't reproduce faithfully.
From the transcript
“I build these GPTs that kind of think like me and the purpose of that is so that my team can get feedback that is…”
“both the speed of those loops and the like the number of those loops that you're able to get is just like radically different with…”
“create ELSAT style questions uh to to test logical reasoning but put them in the scenarios of things that a PM would encounter”
From the episode
How to build a team that can “take a punch”: A playbook for building resilient, high-performing teams
Hilary Gridley (Head of Core Product, Whoop)