𝗪𝗵𝗲𝗿𝗲 𝗗𝗶𝗱 𝘁𝗵𝗲 𝗞𝗻𝗶𝗳𝗲 𝗖𝗼𝗺𝗲 𝗙𝗿𝗼𝗺
In third grade, I wrote a guide on how to make a peanut butter and jelly sandwich.
I wrote four steps:
- Get the bread.
- Get the peanut butter and jelly.
- Spread the ingredients on the bread.
- Put them together.
My teacher asked one question: Where did the knife come from?
I had the vision in my head. I saw the knife in the kitchen drawer that morning. I assumed the reader saw it too. I was wrong.
I made the same mistake when I started prompt engineering. Most people follow this same path:
Phase one: The one-liner. You write one sentence and expect the model to read your mind. When it fails, you fight with the model instead of fixing the prompt.
Phase two: The notebook. You save prompts that work. You notice that consistency matters. You see that some prompts fail because the setup is missing.
Phase three: The environment. You realize a prompt is not just an instruction. A prompt is a room. The model only uses what is in that room. If the knife is not in the room, the sandwich never gets made.
To build better prompts, stack these three things:
- Context: Tell the model what data it has access to.
- Constraints: Tell the model how to use tools and what to avoid.
- Acceptance criteria: Define what a finished result looks like so the model can check its own work.
Forget magic words or clever tricks. Prompt engineering is like writing a bug report or a design doc. Assume the reader lacks your context. Put the context in your text.
Source: https://dev.to/rshade/where-did-the-knife-come-from-1ebp
Optional learning community: https://t.me/GyaanSetuAi