This is Lesson #04 of the ANTIghostwriter course — a free, complete system for creating authentic content with AI assistance.
New here? Start from the full course overview.
Previous lesson: #03: Set Up ChatGPT and Claude System Prompts for Better Output
What You’ll Learn
AI isn’t Google — one-line queries don’t work. In this lesson, you’ll learn how to structure prompts that get consistently great results: using bullet points, setting constraints, defining output formats, and organizing with tags. The more detail you provide, the better your outputs. This foundation makes every future lesson more effective.
Time to complete: ~5 minutes to understand the principles
We’ve already set up a system prompt that contains very important elements that need to be considered when creating any query to artificial intelligence.
Most people approach this very simply—they just write a query in one line, as we’re accustomed to doing with Google. Google is a different story: the shorter the query, the better the result. A keyword or phrase is a way to search for pages on the web. If the keyword or phrase matches what’s on the page, it appears in the results.
Artificial intelligence is completely different. It’s better to think of it as a living person who can think and analyze your query. The key point is to give it as many instructions as possible—as precise a task as possible. The more precise the task and the fewer contradictions it contains, the better the result.
For example, if you ask it to generate an image of a blue-red square, it will make it both blue and red at the same time, because for AI, both colors are commands for action. As an example, when I asked it to draw a blue and red square, it made one half blue and the other red. Moreover, it drew it in the style of Malevich’s square—not perfectly even, with small artifacts in the corner.
To create a good query that will be well-received and produce a good answer, you need to clearly indicate where the instructions are.
- Artificial intelligence models understand bullet points well: numbered lists or bullet points. It’s best to frame them with the word “instructions” and write them out point by point—what you should do, what you shouldn’t do.
- Constraints work very well—rules that must be followed or avoided when generating a response.
- Setting response rules works well: specifying in what format to answer.
- Structuring the query works well: using first and second level headings, tags (in angle brackets, like in HTML).
In our system prompt from the previous lesson, you can see how query structuring is implemented.
You can use different structural elements: instructions, context, constraints, output format, and so on. Here, you can approach prompt engineering creatively. But the key point is that the prompt should not be a single phrase or sentence.
There are cases when such a simple approach works: simple questions, for example, “how much is 2 plus 2” or “what is the weather now.” For questions that don’t require deep reasoning, or where the result isn’t particularly important, this approach is fine.
This also works for instructional prompts—for example, when setting up something. I used ChatGPT to set up a file server: I said I needed to set up a server and asked it to explain options and provide step-by-step actions. That was the entire prompt, without complex instructions. ChatGPT provided a step-by-step plan, and I set up the server, sending the result of each step. Short prompts work for such tasks.
However, for consistent results, you need to establish frameworks, set boundaries, and create behavior templates for the model. This requires long prompts that establish boundaries, indicate how to act, how to respond, and how to communicate.
For tasks like deep research, I recommend detailed prompts that deeply specify the research context. Later, it will become clear why my approach turns out to be thorough and comprehensive in the context of articles.
The main point is: use long prompt formats that are well-understood by models. The more details you provide, the better the answer and the closer it comes to your expectations. If something doesn’t match your needs, edit the prompt: add elements or remove those that interfere.
Don’t be afraid of long prompts. Claude and Anthropic recommend providing the entire context in one query, as Claude rereads the context from the beginning with each query. This is a feature of Claude, so it’s better to reset the context and start a new chat.
ChatGPT maintains memory between chats and tries to remember the context of previous conversations, which you can leverage. Claude is a more delicate model—context is stored within a single chat. It might remember something from other chats, but it works better with a new chat and reset context.