What Is a Prompt?
The term prompt sounds technical, but it is extremely practical. A prompt is the text instruction we give to a generative artificial intelligence in order to get a specific, useful, and verifiable result. In other words, the prompt is the bridge between our intent and the AI's response.
If you work with applications such as ChatGPT, Gemini, Microsoft CoPilot, or Claude - including tools that act as an AI lawyer, such as Praven Intelekt - the quality of the prompt determines the quality of the result you get.
What Exactly Is a Prompt?
A prompt is the clear, structured instruction we give to a generative artificial intelligence in order to obtain a specific and repeatable result.
Think of it as a short but precise brief: you state the goal (what you want to happen and why), you provide relevant context (the situation, the data, the constraints), you specify the format of the result you want (tone, style, length, file or structure format), and you define quality criteria (how you'll know the answer is good).
The more precisely these elements are described, the more predictable, useful, and easy to use the result will be - whether you work in marketing, software development, education, law, or customer service.
The reason prompting works is simple. Clear instructions steer the model toward the most likely and most appropriate responses; when you explicitly state the desired audience, constraints, definitions, and goals, you minimize ambiguity, reduce errors, and increase the chance of getting an answer that can be put to use right away.
Tokens and the Context Window: How Does Generative AI Work?
Tokens
When you write to a generative artificial intelligence, it measures and interprets everything in tokens, not in words or sentences. A token is a small linguistic unit that the model "sees" and processes. In practice, generative models work by predicting the next token. Learn more about generative AI models here.
A token can be a whole word, part of a word, a number, a symbol, or punctuation. The reason models work with units smaller than words is so they can understand rare terms as well as different forms of the same word. As a result, a sentence is counted not in words, but in the number of tokens.
Context Window
The context window is the model's short-term memory, measured in tokens. It includes the system instructions, your prompt, the chat history, and the response itself.
It's important to know that when the total volume of information exceeds the limit of the context window, older instructions get cut off or part of the original data is ignored. This can lead to gaps in the response and incomplete information when using general-purpose applications like ChatGPT. This is also why structured, clear prompts work better.
In addition, it's good practice to keep key facts at the beginning and end of the prompt, since models sometimes handle information buried in the middle of a very long context less reliably.
What Does This Mean in Practice?
-
Provide only the relevant context, and work in chunks when dealing with long material.
-
Start new chats when you switch topics or when you want to attach more documents.
-
Briefly restate the most important parameters of the task in a new prompt, so you don't rely on old instructions that may be ignored.
-
Clearly label the sections in documents you upload, and refer to them by name to make it easier to retrieve information.
-
Set limits on the length of the result (e.g., 1,000 words), so there's room left for the input data and the instructions you've given.
In ChatGPT, the context window of GPT-5 depends on the plan and mode. At the moment, for the individual modes this is approximately 16,000 (Free), 32,000 (Plus), and 128,000 (Pro/Enterprise) tokens.
If you use the GPT-5 API - that is, the dedicated version for developers building their own AI applications - the context window is considerably larger, reaching up to 400,000 tokens.
Quick Tips for a Better Prompt
-
Start with a clear role and the goal you want to achieve.
-
Provide minimal but sufficient context: facts, dates, constraints.
-
Specify the format of the result you want: table, list, summary, code, JSON.
-
Set quality criteria: length, tone, pacing, mandatory checks.
-
Ask about the unknowns. Have the model identify the missing data and state its assumptions.
-
Give examples. A single short example calibrates style and structure.
-
Work iteratively. First produce a framework and a plan, then the details of each part (for example, when drafting or reviewing contracts).
-
Use constraints. Up to 5 bullet points, no more than 150 words, 3 alternatives.
-
Test variations. Small changes in wording often produce a significantly better result.
-
Be consistent with terminology. Use the same names for objects and parts of documents.
-
End with a self-check task. Ask the model to flag risks, ambiguities, and next steps.
You can find more practical guidance on how to create effective prompts in our article here.
At Praven Intelekt, we've already taken care of setting up complex legal prompts - try it out on your own tasks related to Bulgarian law today:
