Most bad AI answers are missing-context answers. The model wrote a good reply for a business that is not yours.
What context means
Context is everything the model can read when it answers. That is your message, the files you attached, your saved instructions and the chat so far.
It cannot read your inbox. It cannot read your head. It does not know what you decided last year unless you tell it.
The one-page context file
Write it once and keep it under a page. Five parts are enough.
- What you sell, in one sentence, with the price.
- Who buys it, and what they were doing before they found you.
- How you sound. Three things you always do and three you never do.
- Decisions you already made, so the model stops reopening them.
- What it must ask you first. Discounts, refunds and anything legal.
Use plain sentences. A model reads this page the way a new freelancer would. If a stranger could not act on a line, rewrite it.
The same request, with and without context
Without context: Write a follow-up email to a lead who went quiet.
With context: I sell bookkeeping for freelance designers at 90 dollars a month. I write short and plain, with no exclamation marks. Write a follow-up to a lead who asked about the price and then went quiet.
The first request gets you an email any company could send. The second gets you a draft you can edit in a minute.
Where to put it
- Custom instructions. The tool reads them in every chat. They have a length limit, so keep only the lasting basics there.
- Projects. ChatGPT and Claude both let you keep files and instructions for one area of work. Good for a job you repeat, such as outreach.
- Pasting. It works in any tool. It is slower, but you see exactly what the model got.
Keep it short
More context does not mean better answers. A model has a limited amount of attention, and every extra paragraph takes some of it.
Anthropic's engineering team describes the goal as the smallest set of information that gets the result. For you, that means one page, and only the parts each task needs. The idea has a name now, and there is a plain explanation of context engineering on this blog.
Keep it current
Put a date at the top of the file. Change it the day a price or a policy changes.
One wrong line gets copied into every draft. The model will not question it. That is the job you keep.
Where a file stops working
A file works well while you have a few tasks and one tool. Then you have ten tasks and four tools, and each holds its own copy. The copies drift apart, and the oldest one always turns up in the worst place.
Nive is being built to remove that step. You tell it how you work once, and each job reads from the same place. You still review everything before it goes out. Join the waitlist to get the demo first.