Summary

  • Allow 'I don't know', require citations, and use direct quotes to cut Claude's hallucinations.
  • A Redditor spotted them in Anthropic's official docs, and these simple prompts improve factual accuracy.
  • The instructions do cut creativity rates, so it's best to make a toggle that activates and deactivates them.

The good people here at XDA have been falling in love with Claude, and we've been on the hunt for ways to prompt it better. Giving Claude instructions is a great way to get it to work the way you want, and while they often require a bit of tweaking to get them just the way you want, that's part of the fun.

Someone on Reddit claimed that there are three commands which, when used, "dramatically" reduces halluciations. The best part is, these are actually listed within Anthropic's documents; it's just that nobody had noticed them until now.

A Redditor has reportedly found the ingredients for fewer hallucinations

And they were under our noses the entire time

Over on the ClaudeAI subreddit, user ColdPlankton9273 stumbled upon a document in the Claude API Docs. Titled "Reduce hallucinations," the document details some simple ways you can prompt Claude to get it to hallucinate less and give better results.

It's worth checking out the page and the Reddit post for more details as to how these work, but the core ideas are:

  1. "Allow Claude to say I don't know"
  2. "Verify with citations"
  3. "Use direct quotes for factual grounding"

ColdPlankton9273 gave these commands a go, and while they did reportedly work for them, it does require a little more tweaking if you intend to use Claude for other tasks (and yes, it's good for more than just coding):

There's a tradeoff though. A paper (arXiv 2307.02185) found that citation constraints reduce creative output. So I don't run these all the time. I built a toggle: research mode activates all three, default mode lets Claude think freely.

One Redditor, Mean_Smell_6469, states that they use Claude as a customer support agent. Before, they state that "Claude would confidently answer questions that weren't in the FAQ at all — just plausible-sounding fiction," but once they added the instructions to ""Allow Claude to say I don't know" + explicit instruction to only answer from the provided FAQ," Claude stopped making things up and began connecting people to the owner. As such, if you've had Claude hallucination issues recently, it's worth giving these rules a try.