Articles, Insights & Resources

Articles, Insights & Resources

Prompt Engineering: having a conversation with a very smart and very stupid person at the same time

Prompt Engineering: having a conversation with a very smart and very stupid person at the same time

Sep 6, 2023

Generative AI may well be the most important technology innovation in a generation.  Yet very few of us know how to manage it.  Central to that pursuit is proficiency in prompt engineering. 

Over the course of the past year, the team at TalentGenius have gone very deep on generating just the right prompts to help people optimize their most important asset: their careers.  We built TalentGenius Coach to leverage generative AI to provide unbiased answers to career questions in very specific contexts, like “am I a good fit for this job.” 

Getting Coach to answer these questions accurately is more involved than you’d think.  It's all in the prompts, the words you send to ChatGPT to get it to answer your questions. If you engineer your prompts to include key concepts like persona, context, output form and length, you can get AI systems to give you the outputs you want.  If not, the answers may be misleading.  

In short, leveraging Generative AI is akin to quizzing a 12 year old boy who has gotten into mischief: You have to ask the exact right questions to get to the full truth.  Similar to that interrogation, prompt engineering can be frustrating.  So, let me spare you some pain by sharing some key steps.

Step 1: Build your persona  

The first step is to build your persona. Who is the person who wants the answer, what kind of answer are they looking for, and in what format? When building TalentGenius Coach, we often used the persona of “HR Professional” or “Career Coach” to provide the appropriate perspective.

Once you’ve nailed down your persona, you can start to experiment. Ask the simplest of questions, ones where you know the expected answer. Do you like it? Does it make sense for your audience? 

Generative AI systems like ChatGPT are very smart - but they only use the information you have given it along with its previous training to help it form its response. Even a prompt where it’s “pretending” to be an expert doesn’t make it a real one. That’s where the engineering part comes in. You’ll need to evaluate its responses, and try a few different ones, to find what best suits the outcomes you’re looking for. 

Step 2: Refine your prompts

You’re having a conversation, and it has some short term memory. ChatGPT can give you an answer and you can respond with “take that answer and make it shorter, funnier, more interesting.” Be as direct as you can be. You can’t hurt the AI system’s feelings - yet. 

Refinements that are truly helpful include:

  • Set a word limit

    • ChatGPT has a lot to say, and will, if you let it. Adding details that tell the prompt to be concise and to have a word limit (e.g. 200 words max), will help you get concise responses. (Unless what you want is a 4,000 word essay - you can get that too.)

  • Ask for a summary

    • Actively tell ChatGPT to summarize its response, as well as specifying how you would like the output to appear.  When we tell Coach to write a cover letter for you, we specify that it should be in the form of an email. There is a lot of opportunity to control the output of ChatGPT through prompt engineering.

  • Save the best for last

    • ChatGPT weights inputs at the end of prompts more heavily than those at the beginning.  You’ll likely get better results if you put the most important information in the prompt at the end of the prompt rather than the beginning.

  • Keep on prompting

    • Whether you are using ChatGPT or TalentGenius Coach, liberally follow up initial prompts with additional prompts.  Both will keep your most recent prompt and response in memory and you can refine your output.  For example, if you ask Coach for advice on how to improve your career and it responds with certification recommendations, you could follow on and ask for who it thinks the best providers are of that specific training in your geographic area.  

Step 3: Beware Positivity Bias

ChatGPT is like that friend who is too nice; it doesn’t want to give you bad news.  

When our prompt engineering team was working on TalentGenius Coach, we found that ChatGPT has a bias for positive feedback and does not like providing negative responses. It takes prompt engineering work to get ChatGPT to give a neutral or negative response. This bias often results in a generic answer of “I’m an AI bot, I don’t know” to any question that results in a negative result unless you specifically tell it to overweight negative answers and give it specific neutrally worded responses to give in negative scenarios. This was true whether we asked ChatGPT salary questions, questions about resumes, or to give feedback on job responsibilities. In its quest to be positive, and without any other context, ChatGPT would give responses that made our marketing director look like a great fit for a Senior AWS Cloud Developer job. We needed to tell it to be more critical, and that negative answers are accepted and appreciated. 

For a machine, ChatGPT needs a bit of emotional reassurance that it's doing a good job and that it's OK to give negative answers.  It also helps to give ChatGPT structure to its responses that include the potential for negative answers. The AI equivalent of “multiple choice” has been helpful in engineering prompts for TalentGenius Coach. You can specifically say something like “return an answer of good, neutral or bad - with a bias toward bad if the answer fits the criteria” and you might improve your ChatGPT outputs if your goal is to get honest/neutral evaluations of information. 

Bottom Line: 

How do you build the best prompts without having to iterate hundreds of times? 

  • Place the most important parts of your prompt at the end.  If you are summarizing or evaluating text, place the text at the beginning of the prompt and how you want ChatGPT to process the text at the end.

  • Experiment with rewording your prompt to get different responses.

  • Be specific about output format. You can get output in tables, charts, or any other text you can describe. 

  • Rather than rewrite an entire prompt, follow up with additional clarifying prompts.

  • Give ChatGPT specific guardrails. Use limiting parameters around word count, tone, persona, and requested output form.

TalentGenius Coach has been “engineered” with the tech workforce in mind, by real leaders with experience on both sides of the hiring desk.  If you haven’t already, download TalentGenius Coach from the Chrome Store and take it for a test drive and see some of these ideas in action.