Tips and tricks

One sentence that instantly improves any Claude conversation, borrowed from how GANs work

The single most useful thing I’ve learned isn’t a prompt template

It’s one sentence you can drop into any conversation at any point:

Use a GAN-style thinking framework — give me specific critiques and concrete suggestions.

When Claude feels too agreeable or surface-level, drop that line. It shifts from “helpful assistant” mode to actually pressure-testing whatever you’re discussing.

Quick background if you’re not familiar: in a GAN, a Generator creates output and a Discriminator tears it apart. The tension between them is what produces quality. You’re telling Claude to stop being a yes-man and start being a sparring partner.

What this actually looks like

I was trying to decide whether to buy a Mac Mini as a 24/7 AI workstation or just rent cloud GPU time. Claude gave me the usual “both have pros and cons” answer. Useless. So I dropped the line.

Claude split into two roles, one going all-in on buying, the other going all-in on renting, and each side attacked the other’s assumptions. The answer I got back was actually useful: “Buy if your workflow is Claude Code + API calls. The Mac Mini isn’t the AI, it’s the cockpit. Rent if you need 70B+ inference locally. And give yourself a kill criteria: if after 2 months you’re not using always-on capability, sell while resale value is still high.”

That “cockpit, not GPU farm” reframe came entirely from adversarial tension. A flat pros-and-cons list wouldn’t get there.

“Isn’t this just pros and cons?”

No. Pros and cons gives you a flat list with equal weight and no judgment. Everything looks roughly equivalent and you’re left where you started. The GAN framework forces each side to actively attack the other’s arguments until something breaks and reforms into something sharper. You get a winner, or at least you get clarity on what actually matters.

This works especially well in Claude’s Plan Mode, by the way. Claude seems more willing to commit to extreme positions there instead of hedging. Worth trying.

Where I keep using this

Architecture decisions. Code review (Claude GANs its own code, which is kind of funny). Writing, when I need to find weak arguments I’m glossing over. And honestly, any moment Claude feels too agreeable, that’s your signal.

One sentence. Go try it.