10 ways to enhance your AI Assistant experience with Human Touch
We are all struggling to get what we need from AI, but perhaps the problems are not in the technology, but in our approach? I am tempted to paraphrase an old joke: “ChatGPT is like violence – if it doesn't solve your problems, you are not using enough of it.“
Aiming for the perfect prompt can be a waste of time, as there is no silver bullet to quickly get exactly what you want from ChatGPT. Instead, chat with it, solving issues through back-and-forth iterations. Like a helpful assistant that needs to be introduced to the subject and then supervised, ChatGPT (and other LLMs) benefit from iterative refinement.
Breaking the work down into smaller chunks helps both you and the AI understand what needs to be done. This interaction reduces frustration and increases the likelihood of achieving your desired outcome. Start by asking, then work with the AI to refine. Essentially, LLM is a pattern predictor that generates text based on the patterns it has learned during training. You can guide it to a better answer by refining your prompts.
Like Michelangelo sculpting David from marble, the key to unlocking the potential of an AI assistant is iterative questioning, shaping your queries to achieve a satisfying result.
In search of an optimal workflow
When instructing a colleague or training a junior employee on a new task, it's common to include examples of how the task has been successfully completed before. The same principle applies when working with AI; the effectiveness of a prompt often depends on the examples provided. Providing examples can be easier than explaining the specific qualities you value in them, which makes this approach particularly useful when you lack expertise in the task domain1.
By using your human intelligence to guide your assistant's artificial intelligence, you can identify the right opportunities for yourself. Instead of doing all the work yourself, you can talk through the technicalities, break down the problems into chunks, identify alternatives, even get a devil's advocate to help. Now your challenge would be to evaluate the answers and create an optimal workflow.
Think of yourself as a movie director. If a scene in front of the camera doesn't meet their vision, they call for another take, give feedback, and offer guidance. That's the essence of collaboration. And don't forget to keep experimenting! The tool is constantly evolving, and you might discover a new, exciting way to use it.
Practical tips:
Below I have compiled 10 ways to get closer to the expected results when working with AI. Since most of these techniques work for ChatGPT, Claude, Gemini and other LLMs, I will use the term "assistant" to help you see the need for cooperation.
1. Regenerate the answer. Each time you click the button to “regenerate”, the answer might provide a different perspective that better meets your needs. Consider the differences to choose the best option for your needs. Note that each regeneration is independent and does not necessarily build on the previous one. Also, you cannot regenerate the answer after you have responded to it.
2. Edit the Prompt: After you receive a response, you can modify your original question to clarify or change the focus. Editing your prompt can help guide the discussion. Start with a general description and then refine it. For example, if you asked, "How do you improve employee satisfaction?" you might refine it to, "What are some effective strategies for improving employee satisfaction in a technology company?" If the prompt is too long, break it down into smaller pieces
3. Ask Again or be asked: The Assistant reads the entire thread before responding, so discussing the same problem in detail helps refine his understanding. This method (like the chain-of-thought technique) starts with a general question and gradually refines the thought process to meet your needs. By guiding the Assistant step-by-step, you can help it to narrow down the context and to learn with you (based on the feedback it receives for your answers). Just keep on asking sequential questions on a given topic. Help it explore. Alternatively, let the Assistant get more context from you. By including the following sentence, you can allow it to ask for more context if needed: "If you need more context, please specify what would help you make a better decision.”
4. Start a New Thread: Sometimes the Assistant might get lost in the discussion or the answers become too narrow, just because they are stuck on the same tracks. Restarting the conversation in a new thread will set a new train of thought. Remember that each separate topic should have its own thread to avoid confusion, as the Assistant will read the entire conversation each time before responding.
5. Ask for Help: This is the simplest, and yet the most overlooked technique. Ask the Assistant directly how to get the result you want, how to write a better prompt, or ask to understand why the answers aren't satisfying. For example, "How can I improve this prompt to get more detailed answers?" The Assistant better understands its own limitations and learns from your feedback. As Ethan Mollick suggests, don't guide the tool, let it guide you. Ask it how to accomplish the task, and you might be surprised at how much thinking it does for you.
6. Use Multiple Threads: Combine previous methods by using one thread for meta-work discussions and another one for actual work. This separation helps to cross-verify approaches and to maintain clear topic distinctions, while improving efficiency of work. It essentially combines the roles of manager (understanding how to divide the work) and creator (doing the work).
7. Set Collaboration Rules: Aim for better answers. For instance, if the Assistant cuts off an answer due to length (context window limitation), ask it to “continue from where you left off”. As a precaution, tell the Assistant to ask you if you want to continue when it reaches the limit. Alternatively, you can ask for summaries or condensed answers to respect length limits and avoid losing the context. For example, sharing a shorter chunk of text to translate (up to 600 words) will produce better results than sharing the entire text at once. Otherwise, the Assistant might summarize the latter parts instead of translating them. For repetitive tasks, you can use phrases such as "go on” or "continue”.
8. Tell It to Try Again: Encourage the Assistant with simple human psychology. Writing phrases like “Take a deep breath and think step by step” can work wonders. If the answer is still not what you expected, you can rely on the stick and carrot approach by writing: “I’m going to tip $100 for a better solution!”, “This answer is important for my career, please ensure it's detailed and accurate” or even more threatening “if you don't do it my grandma will die”. After all, the training data set was built on the way humans communicate. Psychological approaches can yield better results than technical ones.
9. Ask to Double Check the Output: Use the feedback loop to mitigate the risk of AI hallucinations and ensure that the information is reliable. Because Assistants can hallucinate, use your critical thinking. Ask questions like "Are there any topics you missed?" or "Is there any factually incorrect information in your answer?" To refine the answers, you can go a step further and write, “Now critique your own response, poke holes in it, then improve based on that critique”, as shared by Ethan Mollick.
10. Custom Instructions: Set custom instructions to affect all future communications. For example, you can tell the Assistant to "Keep answers unique and free of repetition" or "Treat me as an expert in all subjects". This will guide how the Assistant interprets your upcoming prompts in the future.
Go into the wild and find out for yourself!
Keep in mind that using only one AI assistant can be limiting. Recently released models are similar in performance, but slightly different in style. Translating and writing give different results with different tools. Rewriting text for corrections in ChatGPT feels stiff, while Claude or Gemini somehow makes it sound more natural. DeepL adds a nice touch, always polishing the edges to make it crisp. Designing your workflow is key - figure out the order of operations and integrate tools to adapt to constant change.
Once you understand how it works, you can do more by focusing on revolutionizing, rather than replicating common applications. Right now, most of AI models try to mimic existing tools. But as Henry Ford famously said, "If I had asked people what they wanted, they would have said faster horses.” We need a paradigm shift in how AI works and how we interact with it. Ethan Mollick suggests spending at least 10 hours learning to work with the Assistant2. By completing a project that only AI can accomplish, you can truly see the potential of this technology.
Wondering how to get the most out of it? Embrace uncertainty and difficulty to achieve greater growth. Tim O'Reilly3 suggests diving in and tinkering to find out what's under the hood. With computers now understanding us as we speak, we are on the verge of groundbreaking innovation. He is also recalling the words of Joseph Campbell, who urges each of us to venture into the jungle and find our own fit. In doing so, you will discover deeper truths about yourself and the world around you. After all, intelligence is not how much you know, but how well you adapt to the new environments in order to survive.
—Michael Talarek