Raising an AI Is a Lot Like Raising a Child
The first instructions shape everything, and what you withhold, sequence, or delete matters just as much.
Lightspeed Venture Partners · From Prompt to Loop Engineering | LightworkNomi, an engineer who builds systems for AI agents to work on their own over long stretches, said a prompt is like “the first words that are spoken to the baby.”
A prompt is "a collection of words sent to a model." It is the first information the model receives that it was not trained on, since models learn from vast collections of material from across the internet and then wake up one day to something new. Because those words come first, they shape how it behaves as it goes about its work.
These models are striking because they speak our language. The natural place for people to start was the words they typed into the box. That work still matters. Prompt engineering is not dead, in his phrase, “long live prompt engineering.”
What came next was context engineering. By context he means everything the model can view and use while it works, including its own past output, messages from people, and programmatic responses from the tools and software around it.
For that stage he offered another family comparison. As a baby grows, you might tell it small untrue stories as it develops, like saying Santa Claus is real and later explaining that he is not.
Builders do something similar. The job is not to stuff every fact about the world into view. It is to select the parts most relevant to the task at hand, share them at the right moment and in a helpful order, and sometimes even erase material provided earlier, so the model stays focused on the task.
Questions people ask
What is the context window?
It is the amount of text and other material a model can consider at one time, including instructions, conversation history, tool results, and its own drafts. Common limits today run in the hundreds of thousands of tokens, with a token roughly equal to part of a word, and larger windows cost more and can make models less reliable.
What is the difference between a prompt and context?
A prompt is what the person directly asks or tells the model to do. Context is the broader bundle the model actually sees while working, including the prompt plus prior conversation, retrieved documents, tool outputs, and system instructions added behind the scenes by software.
Why would you ever delete information from a model?
Because models can get distracted, confused, or less capable as their context fills with old drafts, errors, and irrelevant detail. Builders often summarize progress, keep only the most useful facts, and remove outdated steps so the model stays focused on the current goal.