Prompting tips: How a few words can change your gen AI game

A few words can make a big difference when you’re prompting a large language model chatbot.

If you’ve spent time working with an LLM, whether for work or personal use, you’ve probably done a lot of trial and error. Generative AI is neither a one step tool nor something that requires complex codes. Instead, at least at this stage of the technology, it requires problem solving on the part of the user.

Recently I was using a gen AI tool to help find resources using some very specific parameters. It was giving me dead links and even telling me that what I was looking for didn’t exist, which I knew wasn’t true. 

I noticed that the more I told it to redo the request, the longer it would take. I had it set on “thinking mode,” which across the various models is a setting that gives the the system more time to reason before answering, but it still tried to deliver the results as fast as possible. It only took extra time after being told repeatedly that the results it gave were not satisfactory.

The next time I asked it to perform a similar task, I specifically told it to go slow. That accuracy was more important than speed. Six minutes later, it gave me a near-perfect result. 

Game changer.

It’s not foolproof, but it’s one of several small prompt additions that can make a big difference. We asked members of Technical.ly’s community Slack for their own small tips that made a big difference. 

Here are a few of our favorites.

(Not yet in the Technical.ly Slack? Request an invite.)

‘Rate my work’

+

If you want feedback that doesn’t just glaze you, ask the chatbot to score your work on a scale of 0-10, with 0 being the worst possible and 10 being the best possible at achieving your goal. Plus, the LLM will usually explain the reasons for its rating, which can help you pinpoint weaknesses.

‘Make this 100x better’

Chatbots are notorious for delivering results that are unclear, inaccurate or just not quite right. Asking it to make it 100 times better can actually produce an improved response.

‘I will accept nothing but the best’

The full prompt suggested by a Slack member: What changes would you make if I told you I have an extremely high bar for quality and will accept nothing but the best? Explain your reasoning.

“I’ve been really blown away in a good way by the results,” the commenter said.

‘Give me three takes’

Sometimes, the game changing prompt is more than a few words:

• Answer with temperature 0 — be literal and strictly factual.
• Use a temperature around .5 — thoughtful, but grounded.
• Respond at temperature 1.0 — bold ideas, no safe answers.
• Give me three versions at low, medium, and high temperature for comparison.

I tried this out with a simple topic: Should I buy my new favorite album on CD, vinyl or just stream it?

Temperature 0 listed pros and cons of each, and said there was no objectively best format, choose based on your priorities.

Temperature .5 said that streaming is the most efficient, but if the album is really meaningful, CD is the most practical physical format. It suggested using streaming to find music, and buying CDs to keep music.

Temperature 1.0 implored, “Don’t leave it trapped in an app!” It recommended buying the CD, unless I’m really obsessed and already have a turntable, then it may be worth it to get the vinyl.

It’s an interesting approach, and if you know the “temperature” you want for a specific response, it can be a handy addition to a prompt.

Tell me what you got wrong

This one isn’t from our Slack community; it’s a known LLM hack that I first heard from Australian Youtuber Ann Reardon. 

After the chatbot gives you its answer to your query, respond with “OK, now tell me what you got wrong.” It will fact check itself and update its own answer to be more accurate (of course, there is no guarantee that even that will be 100% accurate).

‘What’s good?’

Out of curiosity, I asked the gen AI tool I use most often what it would suggest for a game-changing prompt.

It responded: 

If I had to boil it down to one single best thing to add to a prompt, it would be this: 

Tell me what ‘good’ looks like to you. 

Everything else is secondary. 

Most people ask what they want. The people who get great responses also explain how they’ll judge the answer.

Example: “Help me write an email about X. A good answer would be concise, slightly conversational, and suitable for a business colleague.”

Thanks to our Slack community for the discussion. We chat about AI, tech jobs, the ecosystems and more every day. It’s always free to join.