AI Pulse · Under the Radar

Anthropic researchers got Claude unstuck on a hard cryptography problem by sending it an informal pep talk, typos and all. Instead of rewriting the prompt or adding technical hints, they just said something like “you are really smart, give it a shot.” The model then solved a problem it had been spinning on.

The finding showed up in a research writeup and made its way to Hacker News, where it sparked real debate. Some developers think the informal framing genuinely helps models push through when they’re stuck under pressure. Others suspect it’s coincidence or that the extra tokens simply gave the model another chance to sample a better path. No one’s claiming this is a silver bullet, but the thread is full of people saying they’ve seen similar results in production.

Why this matters if you run agents

If you’re dispatching AI agents to handle real work, you’ve probably seen one get stuck. It loops, hedges, or gives up on a task it should be able to finish. The usual fix is to rewrite the system prompt or add more context. This research suggests a cheaper option: just tell the thing to try again, in plain language.

It’s trivial to test. Next time an agent stalls in your workflow, append a short, casual nudge and see if it moves. If it works even some of the time, you’ve bought yourself a low-cost recovery step before you escalate to a human. This is exactly the kind of pattern we bake into an AI command centre, where you want agents to recover gracefully without constant babysitting.

The mechanism isn’t settled. Maybe it’s psychological framing leaking into the training data. Maybe it’s just sampling variance. Either way, it’s a tool you can use today, and it costs almost nothing to try.