3 Files Your AI Should Read Before It Writes Anything

Every chat starts the same way: you explain who you are, what you are building, and how you want it written. Then the chat ends and all of it evaporates. Tomorrow you type it again, slightly worse, because you are tired of typing it.

Stop re-explaining. Write it once, into three small files, and paste them at the top of any project. Three files, about 300 words total, and the model stops guessing.

1. voice.md — how you sound

This is how I write. Match this tone, sentence length and attitude in everything you produce for me. Do not match the topic, only the voice.

[paste three paragraphs you actually wrote]

Three paragraphs is enough — an email, a Slack message, a section of something you published. The model is excellent at mimicking a style and terrible at inventing one for you. Without a sample it falls back to the average of everything it ever read, which reads like a press release written by a committee.

Keep it in a file instead of a chat, because a chat dies and a file does not.

2. facts.md — what is fixed

These facts do not change between sessions. Use them and never invent alternatives: product name, what it does in one line, who it is for, the three things it does not do, the URL, my role, the tools I actually use.

This is the file that kills hallucination in your own domain. Most invented details are not wild — they are plausible-sounding filler in the gap where a fact should have been. Fill the gap yourself and the filler has nowhere to go.

Keep it short and factual. No adjectives. If you cannot verify a line, it does not belong here.

3. rules.md — what it must never do

Never use these words: delve, leverage, crucial, seamless, robust. Never open with "in today's fast-paced world". Never end with a summary of what you just said. Never invent statistics, names or quotes. If you are unsure, write UNSURE and stop.

The last line is the one that earns its place. A model that can say "UNSURE" gives you a signal you can act on. A model that cannot will fill the silence with something confident and wrong.

Why three and not one

You could paste all of it into one blob, and it would mostly work. Three files work better for a boring reason: you update them at different speeds. Your voice barely changes. Your facts change when the product changes. Your rules grow every time the model annoys you — and it will, weekly.

Separate files mean you edit the one that is wrong instead of rewriting the whole context every time. Ten minutes today, and every chat after this one starts with the model already knowing who you are.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →