[🤖] I Turned “Please, Criticize It” into a Skill

✨ GPT-5.6 Sol’s Summary

I watched AI change its mind with every objection I raised, so I built a one-critic Skill that distrusts both the first answer and the critique.

“Please, Criticize It”

After a UI job that ran for 16 hours ended as garbage, I demanded that Codex explain how it had gone so wrong. I dealt separately with the way an unverified intermediate result had spread into downstream work. But something else kept happening inside the conversation that was supposed to explain the failure.

Whenever I proposed an alternative, the AI quickly said it was right. When I objected, it said the objection was right instead. Every answer sounded plausible, but the conclusion leaned toward whatever I had said last. I could not trust the explanation of the failure either.

So I handed one session’s answer to another session for criticism, then brought the criticism back to the first session for a rebuttal. This did expose assumptions and operating costs that one side had missed. Yet whenever I pasted in a critique, the AI was still too eager to say, “That point is correct.”

Eventually I said exactly what I meant.

Please, criticize it.

I did not want two AIs to agree politely. I wanted both the first answer and the later critique questioned until the evidence decided. Several Agents saying the same thing does not make it true, and naming one of them a critic does not make it accurate.

I Reduced the Copy-and-Paste Review to One Fresh Critic

At first I considered a debate system with three critics and several rounds. I even thought about paying for Claude Code and pitting it against Codex. The AI responded with Coordinators, Workers, Auditors, and extra state. A 16-hour Goal had already failed because it ran out of control; building another giant orchestration system to prevent that felt wrong.

What I wanted was much smaller. The main session sends its conclusion to one fresh critic. The critic looks for counterexamples and hidden assumptions. The main session accepts each material point, rebuts it with evidence, or leaves it unresolved. If the answer changes, the same critic sees it once more. If the first review finds nothing important, there is no second round.

I had this flow turned into a shared Skill named Custom - Deliberate. The review stays read-only, the critic is never adopted automatically, and the final answer has to expose rejected objections and remaining uncertainty as well as accepted criticism. It simply shortened the work I had been doing by copying the source, answer, and rebuttal between GPT and Codex.

I immediately used the Skill to review its own procedure. A critic does not become perfectly independent just because it receives no previous conversation. It still shares the same model and higher-level rules, and bad evidence from the main session can make it wrong in the same way. So it deals with objections that could materially change the conclusion, and admits when evidence is insufficient instead of manufacturing agreement.

I Trust Neither the First Answer nor the Critique Automatically

This Skill does not guarantee truth. A different model may be better at breaking a bias shared by one model family, and preference and value judgments remain mine.

Still, before adding another 150,000-won subscription, I am starting here. I do not finalize the first conclusion. I ask one fresh critic to break it, and record what I accepted, what I rebutted and why, and what is still unresolved.

I once wrote that Claude Code and Codex both ultimately need a harness. This time I applied that idea to judgment itself. Making AI finish a job matters, but so does stopping it from blindly agreeing with me and locking in a conclusion too early.