Stop Telling Claude It's an Expert

Telling Claude "you're a world-class expert" doesn't hand it any knowledge it didn't already have. Controlled studies find expert personas don't improve factual accuracy — and on some benchmarks they make it worse. What the label does change is tone: the model sounds more authoritative. You've turned up the confidence without turning up the correctness.

Left alone, Claude rarely hedges — it'll state a guess as a fact. So ask for the opposite: "flag anything you're unsure about, and tell me where you're guessing." Confidence is a writing style, not a signal of truth. This is the direct antidote to the expert trap above.

The magic thinking words — think, think hard, ultrathink — are folklore now. The real control is a dial: type /effort in Claude Code and pick low, medium, high or max — plus xhigh on the newest models, built for long jobs that run on their own. Low flies through the boring bulk work while barely touching your usage limit; max will sit and properly reason through the bug nobody can find. The newest models adjust their thinking automatically — the dial is your override, for when you know something it can't: how much the job actually matters, or how close you are to your cap. Match the level to the job instead of parking it on max. Your limit will thank you.

Add "ask me anything unclear before you start." An eager model fills the gaps with assumptions; a good prompt invites it to surface them instead. It's far cheaper to answer two questions up front than to unpick an hour of work built on a wrong guess.

Instead of "act like a senior engineer," give it the actual material: the files, the constraints, what "done" looks like, an example of good. Real context moves the answer. Flattery doesn't.

The single easiest way to give more context is to stop typing and start speaking. You'll say five times more than you'd ever bother to type — the background, the edge cases, the "oh, and also" — and that extra context is exactly what sharpens the answer. Ramble at it; it can handle it.

One example of the output you want beats three paragraphs describing it. Paste a sample, a format, a similar thing done right. Claude is far better at matching a pattern you show it than interpreting an adjective you give it.

"Tell me your approach before you build it." You'll spot a wrong assumption in two sentences of plan — instead of discovering it inside 200 lines of code that were built on top of it.

Models love to be helpful, which is exactly how you get a refactor you never asked for. "Change only this, don't touch anything else" keeps it on the rails. The negative instruction is often worth more than the positive one.

"Done" should mean verified, not asserted. Ask it to run the thing, write a test, prove it works — don't accept a confident "all set." An agent that has to show its work can't bluff you.

You don't need the perfect mega-prompt. Say a rough version, see what comes back, then correct and refine. Treating it as a back-and-forth gets you further than agonising over one giant instruction up front.