My favorite adversarial review prompt

Work gets better when it's reviewed after completion. That's at least as true for agentic work as it is for human work.

Many moons ago, Harper taught me a cute trick for getting an agent to think critically about work it had just completed.

Look at this again with fresh eyes.

The phrase "fresh eyes" is kind of magic. You'll usually get startlingly good results with just that one sentence.

But sometimes you need...more.

Having an agent review its own work is a little bit fraught for the same reason that it's not always a good idea to let students grade their own papers or let car companies self-certify their cars' emissions. When you're checking your own work, you have conflicting goals.

And LLMs do really, really poorly when they have conflicting goals.

That's where adversarial review comes in.

At it's simplest, adversarial review is just: Hey Claude, could you please have a subagent check over this work?

Many folks I talk to find that adversarial review works best when you have a different model do the review. Folks who are coding with Claude often use Codex for their adversarial reviews.

Even without a multi-provider setup, I find that I can get really good results by exploiting the models competitive nature.

Please ask two subagents to review this work. Tell them that whomever finds the largest number of serious issues gets five points.

You don't need to tell them what the points are for. It does not matter. The knowledge that they're being evaluated against a competitor seems to do a pretty good job of incentivizing them to work harder. Amusingly, Tell them that whimever finds the largest number of serious issues gets a cookie. seems to work just as well. You can push them a little harder by saying something like I'll be disappointed if they don't find at least 16 significant problems in the work, but I don't usually find that necessary.