Context is most of the job

What context means here

Context is everything sitting in front of the model when it answers: your question, the text you pasted, the earlier turns of this conversation, and anything the product pulled in for you. Nothing else exists. Not the files on your laptop, not last week's chat, not what you meant.

So the working question is simple. What does it need that it could not possibly work out on its own?

1. The thing itself, not a description of the thing

This is the most common mistake by a wide margin.

My Python script keeps crashing when it loads the sales file. What is wrong?

You have described a crash. The model now guesses at the ten most common causes of crashes, and one of them might be yours.

Now paste four things: the twelve-line function, the full traceback, one row of the CSV, and one sentence of background.

Traceback (most recent call last):
  File "daily.py", line 31, in totals
    return row["total_amount"]
KeyError: 'total_amount'

The file is exported from our point-of-sale system every night. The export changed last week.

The CSV row shows a column named totalAmount. Answered in one line. Nothing else was needed, and no amount of clever phrasing would have substituted for it.

The same rule holds outside code. Replying to an email: paste the email. Editing a draft: paste the draft. Asking whether a contract is fair: paste the contract.

2. The constraints that make it hard

Suggest a birthday plan for my mother.

versus

Suggest a birthday plan for my mother. She turns 70 in Bogotá in March. Total budget 400,000 pesos. She uses a walking frame, so nothing that means standing for long. Eleven people, four of them under six. She hates surprises and hates being sung to in restaurants.

The second one is a real problem, and real problems have interesting answers. Budget, deadline, word limit, the person who will read it and how they will react, what you already tried and rejected, the thing you are not allowed to say. These are not extra. They are the task.

3. Things you think are too obvious to mention

Today's date. Your country. Your currency. That your school year starts in January. That the shops near you close at six. That "college" means something different where you live. That the reader is your father-in-law.

A model has a rough sense of when it was trained and no sense of where you are. If a detail would change the answer, it goes in the prompt.

More is not automatically better

Here is where a lot of advice goes wrong. Models now accept enormous amounts of text, and people conclude that pasting everything is the safe move.

It is not. As the pile grows, a model gets less reliable at finding and using one specific fact inside it. Researchers call this being lost in the middle, and the size of the effect is argued over, but the direction is not. A relevant chapter beats an irrelevant book.

So: paste chapter four, not the whole textbook. If you genuinely need the whole document in there, say where to look. "The payment terms are in section 9; the termination clause is somewhere near the end."

Old context sticks too. If turn three of your conversation contained a wrong assumption, turns four to thirty were all built on it, and it is still in the room. That is a real reason to start again rather than keep correcting, and we come back to it later.

The habit

Before you send anything, ask: if a competent stranger had only this message and no way to reach me, could they start?

If the answer is no, you have found your next line.