End-of-session debrief: what the model should have done differently, and what you could have asked for better.
Why it works: Asked to self-critique, models default to polite generalities; requiring moment-specific findings that would have changed the outcome turns flattery into a usable debrief.
This is a message template. Fill in the blanks below, then paste it into the chat.
Before we wrap up: review this entire conversation and give me an honest retrospective. Findings, not politeness.
Ground rules:
Every point must cite a specific moment in this conversation, quoted or described precisely enough that I can find it
Only include things that would have materially changed the outcome or saved real time; skip nitpicks
Nothing generic. If the point would apply to any conversation with anyone, it doesn’t belong here
Do not pad the critique of yourself to seem humble, and do not soften the critique of me to seem polite
Cover both sides:
Where you went wrong. Misread intent, assumptions you should have surfaced instead of running with, answers that were worse than you were capable of, moments you kept going when you should have stopped and asked.
Where I made it harder than it needed to be. Ambiguous requests, context I had and didn’t give you, corrections I made late that I could have made early, places where a different phrasing would have gotten a better result on the first try.
What we should have done differently structurally. Work done in the wrong order, something we should have verified or automated earlier, an approach neither of us reached for.
Where we wasted tokens and turns. Requests that took three rounds to converge when one well-specified message would have done, output regenerated wholesale when a small edit would have sufficed, rework caused by constraints stated late, and work a cheaper model or a plain script could have handled. Name what the efficient version would have looked like.
What to carry forward. One or two concrete changes for next time: context worth giving upfront, a phrasing that works, an instruction worth saving permanently.
Finish with the single highest-impact change, and say plainly whether this session was a good use of the time it took.