Explain like I'm smart, not five.
Assume intelligence, not knowledge. One concrete example, the intuition behind it, and the common misconception you'll immediately hit.
You want to understand something outside your field — a technical concept, an economic term, a biological process — and every explanation is either too basic (ELI5, condescending) or too advanced (textbook, unmotivated). This prompt hits the middle: writes for a smart reader who doesn't happen to know this yet.
THE PROMPT
Explain [CONCEPT] to me. Assume I'm smart, curious, and don't already know this specific thing. My general background is [YOUR BACKGROUND — 1-2 sentences]. Structure the explanation as: **The one-sentence version** The tightest possible statement of what this concept is. If someone stopped reading after this sentence, they'd still have the core. **A concrete example** One specific example that makes the abstract thing feel real. Not a "for instance" that could apply to anything — a specific case with numbers or a specific scenario. This is where most explanations fail. **The intuition** Two paragraphs. Why does this work? What's the mental model that makes it click? Skip the historical detour unless the history is the intuition. **The common misconception** One paragraph. What people who kind-of-know this get wrong. Naming this pre-empts confusion I'd otherwise hit later. **Where to go from here** Two-line "next" — if I want to go deeper, what's the specific next concept I should learn, or what's a specific text/paper that opens the door further. Do not include a "conclusion" or "summary". The explanation ends at the last section. Do not add hedges like "of course, this is a simplification" — every explanation is a simplification; naming it doesn't help.
**The one-sentence version**
A vector embedding is a way to turn a piece of text (or an image, or an audio clip) into a list of numbers such that similar things have similar lists.
**A concrete example**
"The cat sat on the mat" and "A cat is sitting on a mat" would produce two vectors of, say, 1,536 numbers each. If you compare the two vectors mathematically, they'd be very close — much closer than either would be to "Steel prices rose 3% in Q2". You didn't tell the system that the two cat sentences were related; the embedding process figured it out from the pattern of the words.
**The intuition**
Every word (and every combination of words) can be thought of as sitting somewhere in a very high-dimensional space — 1,536 dimensions in the OpenAI text embedding case, or 3,072, or 768 depending on the model. The training process pushes text with similar meanings toward similar coordinates in that space. Once you have coordinates, "similarity" becomes a math problem: how close are these two points?
The reason this is useful: you can now search by meaning, not by keyword. "Find me articles about people leaving their jobs" will match "engineers who quit" without needing the words to overlap. That's the entire foundation of modern semantic search and RAG.
**The common misconception**
People often think embeddings "understand" the text — like the model has an opinion about it. It doesn't. It has a coordinate for it. The coordinate happens to be close to other things a human would call semantically related, which makes it feel like understanding. But there's no meaning inside the embedding vector itself; there's just position.
**Where to go from here**
Read the Sentence-BERT paper (Reimers & Gurevych 2019) for the training approach, or if you want something practical, walk through how a RAG system uses embeddings step by step — that's where the concept really clicks.