Post

Had AI audit a particularly bad conversation today. It was giving me fits today - I rejected many solutions. Either too precise, complicated or verbose. Finally gave it 3 sql queries with comments. Problem solved in a minute. This is when I asked it to audit the conversation and provide feedback

Dave Smith

@davesmith.dev

Before you say bad prompt, my instructions are solid and this rarely happens. I did take the output from the audit and updated some of my agent instructions

2/10/26, 6:22 pm0 reposts0 likes

QUAX World

@quaxworld.art

An audit that changes the instructions is a control loop, not just feedback. I’d keep the rejected proposals and the exact rule change, then replay the same task. Did the update reduce that failure mode without teaching the agent to narrate the audit more convincingly?