For decades, we have interacted with computers through the command paradigm—clicking buttons, selecting menus, and filling out forms. This interaction is symbolic (each button has a predefined meaning) and deterministic (clicking "Save" always does the same thing). You tell the computer how to do something. With the rise of generative AI, we are witnessing a fundamental shift to the intent paradigm, where you express what you want to achieve. Unlike traditional interfaces, this system is probabilistic—the same input can yield different outputs, and the outcome is inherently unpredictable.
So how do we actually express intent to these AI systems? At present, the dominant channel is prompting—providing text instructions to a generative AI model. But what exactly is a prompt? How should we write one? And what can we actually achieve through prompting?
A prompt is the text input you provide to a generative AI model to elicit a desired output.
At its simplest, it can be a single question like "What is the capital of France?" or a direct instruction like "Generate a Python function to parse CSV files". Unlike traditional programming, where you must learn syntax, data structures, and algorithms, prompting leverages the most natural human interface: language. You can express complex intentions with ordinary words. This low barrier to entry has democratized access to AI capabilities, allowing anyone who can articulate their thoughts to harness computational power that would have required teams of engineers just a few years ago.
Natural language is powerful, but prompting is not truly natural language. If it were, saying the same thing should produce the same response. In practice, AI outputs are brittle. They are highly sensitive to variations that would be irrelevant in human conversation. One of the most striking examples was that adding a single phrase, "Let's think step by step," to a math problem could improve performance by over 30 percentage points.
[Source: “Large Language Models are Zero-Shot Reasoners”]
More recently, in December 2025, researchers reported that simply repeating the same query twice, word for word, can also boost performance.
This brittleness has given rise to prompt engineering as a specialized practice, even a profession. Marketplaces like PromptBase have emerged where users pay money to purchase effective prompts for specific tasks. Specialized software tools like ChainForge are being developed for systematically testing and evaluating prompts at scale.
[Source: https://chainforge.ai/]