When one model remains stuck, give a second model the original request, failed output, and desired example, use its critique to identify the mismatch, and test a targeted revision instead of repeating the same prompt.
Source episode · 33:00
How I AI: How Tomasz Tunguz digests 36 weekly podcasts without spending 36 hours listeningwith Claire Vo
When a model keeps producing the wrong result, Tomasz gives another model the original input, the unwanted output, and an example of the result he wants, then lets the models critique and refine the work through a small script.
A revised prompt or output that addresses a diagnosed mismatch and passes fixed acceptance tests, with the model contributions recorded so the result can be reproduced.
Anthropic AI assistant
Google AI assistant
Step by step
Follow the sequence once, then adapt the prompts, checks, and handoffs to your own setup.
4 steps
Save the original prompt, relevant context, model and settings, exact output, and the acceptance criterion it missed. Reduce the complaint to an observable mismatch rather than describing the output as merely bad.
Create a small evidence packet with the original input, failed output, desired example, and several nearby test cases. Remove unrelated or sensitive context before sending it to another model.
Ask a second model to diagnose the failure before it rewrites anything. Require a comparison of the instruction, actual behavior, desired behavior, and the smallest prompt or context change likely to help.
Diagnose this model failure. Original request: [input]. Actual output: [failed output]. Desired behavior: [example or criteria]. Test cases: [cases]. Identify the specific mismatches, likely instruction or context causes, and the smallest correction. Do not rewrite the output yet.
Apply the best diagnosis to a revised prompt, run it with the original and nearby test cases, and compare the results. Keep the competitive framing playful if it helps, but accept the revision only when it passes the fixed criteria.
Using this diagnosis: [diagnosis], produce a revised result for [input]. Meet these acceptance criteria: [criteria]. Do not copy the desired example verbatim. Then report which change addressed each mismatch and how the result performed on [test cases].
Turn an idea into a PRD, user stories, and a plan.
After the steps
How to recover when the loop fails and where human judgment helps.
Recover
Keep building
Design a weekly meta-workflow that analyzes your interactions with your AI assistant to make it smarter. This system learns your writing style from edits, suggests new automations for repetitive tasks, and helps you filter valuable AI techniques from hype.
Create a daily AI workflow that summarizes yesterday's meetings, identifies urgent tasks, and proactively learns internal company terms it doesn't understand. This keeps your AI assistant's knowledge current and highly relevant to your work.
Build a recurring AI task that automatically gathers updates from your calendar, Slack, and notes to suggest weekly priorities and prepare you for upcoming meetings. Start every Monday with a clear, comprehensive plan without the manual effort.
Join 100,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.