So far I have built and shipped a B2C app, an effective system for language learning, I designed a large agentic workflow for a heavy process at work (Fortune 500 company), and I’m currently developing a software that manages a flow of services in a small company. And yet the number of serious prompts (well-described goal, context, extensive instructions) I have written myself so far remains… zero.
Although once a very hot topic, prompt engineering as a concept has gone through a rollercoaster of attention, from the key element of unlocking AI value, through something that requires separate training, to something that current models don’t really need to work well.
How to write prompts
Do you know all elements of good prompt? How to structure it, what kind of instructions to give, how to describe the context or goal? Do you understand how different models react to different levels of control and whether they work well with plenty of freedom or they prefer having clear steps to follow?
Based on above questions, are you 100% convinced you should be the one to write your own prompts?
My answer is no. Actually, I’m convinced the opposite: I always let AI write my prompts for me. The key is to understand how to do it, as this is not as simple as saying “write a prompt for AI to do X and Y”. Adding a few more steps to the process will drastically change the output AI is able to give you based on submitted request.
Key elements of an effective prompt
Goal. Current models are very good at delivering the goal you request once they understand it well. They may know better ways to do it, they have the right tools available and they were trained to give you the outcomes you are asking for. But they have to understand well what you need and half of your effort should go into exactly that.
Context. The quality of AI answers is directly correlated with the amount of personalized information they have about you and things around your request. This is why nowadays models have memory, can read other chats, or AI labs introduced Project-like features where you can submit additional documents with context. Don’t rely only on what is provided, you have to actively build a large context for AI to give you personalized outputs that answer your requests.
What about extensive instructions? Companies like OpenAI or Anthropic realized we are not good at writing prompts. Their models developed ways that help them deliver the goal even with poorly written requests. Heavy set of instructions can actually work against you, as it will be limiting the model in tools/ways it selects to deliver the outcome. This is why you should keep it to a minimum and only add instructions that are relevant for the outcome (for example you want the model to only build the answer based on scientific sources, or you need it to get to an end point in a certain way).
What does it look like in practice
Have a pre-work chat with your model. Say the goal of the conversation is to write a prompt for an AI model to perform the work. Explain the goal you want to achieve later in the main conversation, add as much context as you see necessary and then say:
“Ask me 2 more questions about the goal, 3 questions about the context and 2 questions about the process, so that you have enough info to write the prompt.”
The number of questions is arbitrary – see what works best for you. The part on the process is only important when you think the workload is high (you may ask to perform the work in steps; to review partial outputs; or to have another model validate important elements). If you need it, you can finish your answers with “if you still have gaps, ask additional questions – do not write the prompt until all is clear”. If you are at the stage where you see the difference in models, and you have access to different ones, you can also ask which model will be best to deliver the output for this exact prompt. Apply this approach and see the quality of your work with AI increase.