Ask ten people how much they tell an AI before they ask it something, and you'll get two camps. One types a single line and hopes. The other pastes in everything they can find: the whole brief, three old emails, a résumé, and a paragraph of life story, just in case.
Both camps get disappointing answers. They just get disappointed in different ways.
So here's the practical middle: what goes wrong at each extreme, why more isn't automatically better, and a simple test for deciding what belongs in the prompt.
How much context should you give AI?
Give an AI enough context that it could not mistake you for anyone else, and stop there. For most people that means one or two pages of durable background (role, audience, goals, voice, and current priorities) plus the specific material the task in front of you needs. Anything the AI wouldn't change its answer over can be left out.
More is never the goal. Anthropic's engineering team put it plainly in its guide to context engineering for AI agents: the aim is to find "the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome." That was written for developers building agents, but it holds just as well for one person typing into a chat window.
| Too little | About right | Too much | |
|---|---|---|---|
| What you give it | A one-line request | Durable background plus task material | Everything that might be relevant |
| What comes back | Generic, average-person answers | Specific, usable work | Answers that fixate on side details or miss the point |
| What it feels like | "This could be for anyone" | "This sounds like it knows me" | "Why is it talking about that?" |
| The fix | Add who, for whom, and what good looks like | Keep it current | Cut, and put the key instruction last |
What happens when you give AI too little context?
You get the statistical middle.
A model with no background answers the question the way it would for the average person who has ever asked it. Ask for a marketing email with no context and you get a perfectly competent marketing email for nobody in particular. Ask for career advice and you get advice for a generic professional of no fixed industry, age, or ambition.
The model fills the gaps you left with the most probable answer, which is the definition of generic. This is also why every new chat seems to start from zero: unless something loads your background, there's nothing on the desk.
Too little context is the more common failure by a wide margin. Most people never give an AI the three things it most needs: who you are, who the work is for, and what good looks like.
Can you give AI too much context?
Yes, and the research is fairly consistent about it.
In 2023, researchers at Stanford and elsewhere published Lost in the Middle, which found that language models use information best when it sits at the beginning or end of a long input, and that performance "significantly degrades" when the relevant part is buried in the middle. In July 2025, Chroma tested 18 models, including GPT-4.1, Claude 4, and Gemini 2.5, and reported in Context Rot that "model performance degrades as input length increases, often in surprising and non-uniform ways." Even a single irrelevant but similar-looking passage lowered accuracy.
Bigger context windows don't solve this. A window that holds a whole book tells you how much a model can read. How well it weighs what it reads is a separate question. (The context window vs. memory post covers that difference in full.)
But length is the smaller problem. The bigger one is noise.
Every extra detail is a signal the model might act on. Paste in an old project brief and it may borrow last year's goals. Mention offhand that you once worked in finance and the advice may suddenly lean on finance analogies. Include three sample emails written in three different moods and it has to guess which voice you meant. Stale detail is the worst kind, because the model treats it as current and acts on it with full confidence.
Over a long conversation, the same thing happens slowly. The early setup gets crowded out by everything that came after it, and the answers drift back toward generic.
A simple test: always, this task, or never
Before you paste anything, sort it into one of three buckets.
Always. Things that are true no matter what you ask. What you do, who you serve, how you like to communicate, what you value, and what you're working on this season. This bucket is small, durable, and the same every time, which is exactly why you shouldn't retype it. Write it down once in a short context file and load it at the start of a session.
This task. Material that matters for this request and no other. The draft you want edited. The client's brief. The data for this analysis. Two examples of the tone you want. Include it in full when the task needs it in full, and in a summary sentence when it only needs to know it exists.
Never. Everything else. Credentials the task doesn't need. Background that wouldn't change the answer. Old versions of things. Sensitive details you wouldn't want stored. The test for this bucket is blunt: if the AI would give the same answer without it, leave it out.
Then one placement rule. Put durable background at the top and the actual instruction at the end, closest to where the model starts writing. In a long prompt, the middle is where things go to be ignored.
What does the right amount look like?
Take a hypothetical development director at a small literacy nonprofit who wants help with a grant follow-up email.
Too little: "Write a follow-up email to a funder." What comes back is polite, forgettable, and could have been sent by any organization in the country.
Too much: she pastes the full strategic plan, last year's annual report, her own bio, the foundation's 40-page guidelines, and four old emails. The draft comes back quoting a program the nonprofit retired last spring and opening with her résumé.
About right: her always-on file covers the organization in a paragraph, who she writes to, and her voice (warm, specific, never begging). The task material is the funder's name, the date of the last conversation, the one question the program officer asked, and the two numbers that answer it. Then the instruction, last: write a five-sentence follow-up that answers the question and asks for a call.
That version runs maybe 400 words of context in total. It gets a usable draft on the first try.
How do you know you got the amount wrong?
Read the answer for symptoms.
If it could have been written for anyone, you gave too little. Add who you are, who it's for, and what good looks like.
If it fixates on a side detail, repeats something outdated, or misses the main point, you gave too much, or the key instruction got buried. Cut the extras, start a fresh chat if the current one is long, and move the instruction to the end.
If it was right last month and isn't now, your context has gone stale. A few minutes of upkeep every few weeks fixes most of that.
Where to start
What most people need is the right context, written down once instead of improvised every time. That discipline has a name, context engineering, and the personal version comes down to the always-bucket: a few short documents that tell any AI who you are before you ask it anything.
RUMO breaks that bucket into six focused documents, so you can update the one that changed without rewriting the rest.
Start with values, the part no AI can infer from your files. The free Personal Constitution builder takes about thirty minutes, and it's the first page of the always-bucket.
The one-line camp and the paste-everything camp are both trying to answer the same question. The answer is smaller than the second camp thinks and more specific than the first.
Frequently Asked Questions
- How much context should I give ChatGPT?
- Enough that it could not mistake you for anyone else, and no more. For most people that means one or two pages of durable background, covering role, audience, goals, voice, and current priorities, plus the specific material for the task in front of you. Anything the AI would not change its answer over can be left out.
- Can you give AI too much context?
- Yes. Research on long inputs shows models use information at the start and end of a prompt more reliably than material buried in the middle, and accuracy tends to drop as inputs grow longer and include irrelevant detail. Extra context also adds noise: an outdated fact or a side note can pull the answer in the wrong direction.
- What information should I give AI about myself?
- Start with what never changes from task to task: what you do, who you serve, how you like to communicate, what you value, and what you are working on now. Skip life history, credentials the task does not need, and anything sensitive you would not want stored. Precise, specific sentences beat long general descriptions.
- Why does AI give generic answers?
- Usually because it has too little context. Without background, a model fills the gaps with the most likely answer for an average person asking that question. The fix is to tell it who you are, who the work is for, and what good looks like, so it has something specific to work from instead of a statistical middle.
- Should I paste my whole document into an AI chat?
- Only if the task needs the whole document. For a summary or a full edit, paste it. For a question about one section, paste that section and say what the rest covers in a sentence. Long pastes dilute the parts that matter, and in a long chat the earliest material can eventually fall out of the model's context window.
