Most generic answers come from a request with no situation around it. Brief AI like a capable new colleague, keep only what changes the answer, and save the rest.
Give AI the situation a new colleague would need, then the task. Run every line through one question: would the answer change without it? Describe situations rather than people, so nothing confidential goes in. Save the stable half of your context once, in a personal setting, a project or a file you control, and audit it monthly so it doesn’t steer answers with last quarter’s facts.
Most generic AI answers come from a request that described the task and left out the situation. “Write a summary of this report” tells the model what to do. It says nothing about who will read the summary, what they already know, what decision it feeds, or what went wrong the last time someone sent them one. So the model fills those gaps with the most average reader it can manage, and the answer comes back polished, accurate and aimed at nobody in particular.
That’s the gap context fixes. When I talk about context for AI, I mean everything you give it before the ask: who you are, who the output is for, what it needs to achieve, the limits it has to respect, an example of what good looks like, the material it should work from, and the preferences you’d otherwise repeat every time. The prompt is the instruction. The context is the briefing that lets the instruction land.
Anthropic’s own documentation puts this in workplace terms. It tells people to think of Claude as “a brilliant but new employee who lacks context on your norms and workflows.”[1] I find that the most useful frame for non-technical teams, because everyone already knows how to brief a new colleague. You wouldn’t hand a new starter a one-line task on their first morning and expect the version your manager likes. You’d tell them who it’s for and what tripped people up last time.
Brief AI the way you’d brief a capable new colleague: the situation first, then the task.
This guide covers that briefing and where to keep it. For the wording of the request itself, see our pieces on why prompting matters and reading an answer before you rewrite the prompt.
Context sorts into seven kinds. You won’t need all seven every time; the keep-or-cut test later helps you choose. The kinds people tend to skip are the ones that move the answer most. “Who I am” gets written in almost every time. “Who it’s for” and “what it must not do” get left blank.
| Kind | What it tells AI | Example line (an HR business partner writing a policy update) |
|---|---|---|
| 1. Who you are | Your role, what you own and what you don’t | I’m an HR business partner for an operations division. I own how policy is communicated, not the policy itself. |
| 2. The audience | Who reads the output and how they read it | Shift supervisors who read on their phones between tasks and skim anything longer than a screen. |
| 3. The goal | What should be different after they’ve read it | Supervisors know what changes next month and the one thing they need to do before then. |
| 4. Constraints | Limits, and things only you know | Under 200 words. No legal language. Don’t mention the pay review, which hasn’t been announced. |
| 5. An example | What good looked like last time | Here’s last quarter’s update, which landed well. Match its length and tone. |
| 6. Source material | What to work from, and only from | The approved policy text is pasted below. Work only from it. |
| 7. Standing preferences | How you like every answer, every time | UK spelling. Plain sentences. Bullet the actions, not the explanation. |
A Future Factors working list, not a research-validated model. The right-hand column shows one request filled in so you can see the standard.
Two of these do more work than they look. Constraints carry what only you know, like the announcement that hasn’t happened yet. And an example of past work tells the model more about tone and length than any adjective. Anthropic’s guide calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure.”[1] That fits how it usually goes in practice: “make it friendly” gets you the model’s idea of friendly, and a pasted example gets you yours.
A finance analyst has to explain a budget overspend to a department head who isn’t a finance person. The quick version of the request goes in first, and what comes back is correct and unusable: an explanation of what a variance is, a list of common causes, and a suggestion to review spend regularly. Nothing in it is wrong. It just isn’t about this department, this number or this reader.
| Without context | With context | |
|---|---|---|
| The request | Explain why the marketing budget is over this quarter. | I’m a finance analyst writing to the head of marketing, who is sharp but not a finance person and wants to know whether she needs to act. Goal: she understands the overspend in two minutes and knows the one decision in front of her. Constraints: under 150 words, no jargon, don’t assign blame, and don’t speculate beyond the figures below. Source: the variance table pasted below, which shows events spend running ahead of plan because a conference deposit was paid early, while digital spend is on plan. Style: answer first, then the reason, then the decision. Task: explain why the marketing budget is over this quarter. |
| What it tends to produce | A general explainer on budget variances that could go to any department. | A short note that says the overspend is mostly timing, names the deposit, and asks one question: pull planned spend forward or hold it. |
The task line is nearly identical in both versions. Everything that changed the answer sits in the situation around it.
Notice what the second version didn’t need: a cleverer instruction. The gain comes from a few sentences of situation, and three of them (the reader, the goal and the line about blame) are things the model could never have inferred from a spreadsheet. If you want to see how one request like this can be shaped further across a few turns, our guide to prompting AI like a conversation picks up where this leaves off. “Could the model have guessed this?” is also the first question to ask about any line of context, and it’s where the next section starts.
Once people discover context, the next move is predictable: they paste everything. The whole policy manual, three old email threads, a long biography and a note saying “use whatever’s relevant.” It feels thorough, and it usually makes the answer worse, for two reasons.
The first is that irrelevant material competes with the material that matters. A 2023 study of how language models use long inputs found performance “is often highest when relevant information occurs at the beginning or end” and degrades when the model has to find it in the middle.[2] Models have moved on since then, so I’d hold the details loosely. Anthropic’s current advice to its own developers points the same way, though: find “the smallest possible set of high-signal tokens” for the job.[3]
The second reason is staleness. Old context reads as current. If last year’s org chart or a superseded price list is sitting in a project, the model will usually treat it as confidently as everything else, and nothing in the tone of the answer warns you. Long conversations have their own version of this, which we cover in what is context rot. Here the problem sits one step earlier, in what you choose to include before the conversation starts.
| Question | If yes | If no |
|---|---|---|
| 1. Would the answer change if I left this out? | Keep it | Cut it |
| 2. Is it still true today? | Keep it | Update it or cut it. Out-of-date context usually does more harm than a gap |
| 3. Does the model need all of it? | Paste the whole thing | Paste the part that applies, or summarise it in two lines |
| 4. Is it safe to share under my organisation’s AI policy? | Keep it | Remove it, or describe it without the identifying detail |
A Future Factors working test. Run it on a context block before you save it, and again whenever answers start drifting.
If it wouldn’t change the answer, it doesn’t belong in the context.
Privacy needs its own rule, because context is exactly where sensitive material slips in. A request feels like a request. A context block feels like background, and background is where people paste the client’s name, the salary band, the board paper or a colleague’s performance notes “just so it understands.”
Where that material ends up depends on the tool and the account. OpenAI, for example, says it may use content from its services for individuals to train its models unless you turn off “Improve the model for everyone” in Data controls, and that by default it doesn’t use inputs or outputs from ChatGPT Business, Enterprise, Edu or its API to improve models.[4] Other vendors set their own terms. So the first check is your organisation’s AI policy and which account you’re signed into, before any setting inside the tool.
My own line is stricter than most settings pages. Unless your organisation has approved that specific use in writing, these stay out of the context whatever the plan:
Most context doesn’t need the sensitive detail anyway. “A senior engineer whose last two projects ran late” gives the model the situation. The name adds nothing to the answer and a great deal to the risk. Describe the shape of the situation and keep the identifying details on your side of the screen. For the settings in each tool, see our guide to using AI at work without leaking company data.
Describe the situation, not the person: if a detail identifies someone or something confidential, leave it out or replace it with a description.
Retyping the same background into every chat is how good context habits fade by week two. Write the stable parts once and keep them where you can paste or attach them in seconds. The stable parts are who you are, the audiences you write for most, your standing preferences and your always-and-never lines. The changing parts, the goal and the source material, still go in fresh each time.
| ABOUT ME: I’m an operations manager at a mid-sized logistics company. I own the weekly performance report and supplier communications. I don’t approve spend beyond my budget line; finance does. WHO I USUALLY WRITE FOR: (1) The leadership team: short, numbers first, they want the decision. (2) Site supervisors: practical, reading on phones, actions in bullets. (3) Suppliers: polite, firm, no internal detail. HOW I LIKE ANSWERS: UK spelling. Plain language. Lead with the answer. Flag anything you’re unsure of instead of guessing. If my request is ambiguous, ask me one question before you start. ALWAYS: When I give you a source, work only from it. Tell me if anything I’ve given you looks out of date or contradicts itself. NEVER: Invent figures. Name a supplier or a member of staff unless I’ve named them in this chat. THIS TIME (fill in per request): Goal: ___. Audience: ___. Constraints: ___. Source: ___. Example of good: ___. |
Replace each line with your own. Everything above THIS TIME is stable and worth saving. The last line changes with every request, which is why it’s the only one left blank.
The block shifts by role more than people expect, even though the seven kinds stay the same:
Every major assistant now has somewhere to keep context, and they behave differently enough that it’s worth choosing on purpose. The decision rule I’d use: put what applies to almost everything you do in a personal setting, put what applies to one piece of work in a project, and treat memory as a convenience you check, never as a filing system you rely on.
| Container | Best for | What the vendor says | Watch out for |
|---|---|---|---|
| Personal or custom instructions | Who you are and standing preferences | Account-wide settings that apply across your chats | Keep them short and general. Project settings can override them |
| ChatGPT project instructions and files | One client, one report, one recurring piece of work | Project instructions “apply only within that project and override your global custom instructions”[5] | Your global preferences won’t carry into the project unless you repeat them there |
| Claude project instructions and knowledge | The same, for Claude users | Anything uploaded “will be used across all of your chats within that project”; free accounts are limited to five projects[6] | Upload the current version only. An old draft in project knowledge is stale context |
| Gemini Gems | A repeat task with its own instructions and files | Gems take instructions and knowledge files; Google says that starting in November 2026, Gems “will go away on personal Google accounts” and move to skills[7] | On a personal account, keep a copy of your instructions somewhere you control |
| Memory | Small personal facts and preferences | Saved memories are “details that you explicitly ask ChatGPT to remember”; temporary chats don’t create or update them[8] | You don’t always see what it draws on. Review the list and keep sensitive detail out |
| A saved context file | Anything you use across more than one tool | Works anywhere you can paste or attach text | Only as current as the last time you edited it |
A Future Factors decision table built from each vendor’s help pages. Features change often, so check before you rely on one.
For setup, see our walkthroughs of ChatGPT custom instructions, ChatGPT Projects, Claude Projects, Gems in Gemini and how AI memory works.
Storing context changes who does what. The model applies the briefing; you decide what’s in it and own what leaves your desk.
| What AI does with it | What you still own | How it gets checked |
|---|---|---|
| Uses your context block to pitch tone, length and audience on every draft | What goes into the block and what stays out | The monthly audit below, ten minutes |
| Drafts from the files in a project | Which version of each file is current | Remove a superseded file the day its replacement goes in |
| Recalls preferences you’ve asked it to keep | What it’s allowed to remember | Open the memory list once a month and delete anything stale or sensitive |
| Produces a draft pitched to the situation | Every figure, name and claim that leaves your desk | Check them against the source before you send |
A Future Factors working split for any assistant that stores context.
Context goes stale quietly. A team changes, a priority shifts, the audience you described as “skims on a phone” gets replaced by a board that reads every word, and the block keeps steering the model toward last quarter’s situation. Nothing announces it. The answers drift slightly off, and it takes a while to notice that the cause is the briefing rather than the model.
| Check | Passes if | If it fails |
|---|---|---|
| 1. Role and remit | It still describes your current job and what you own | Rewrite the line |
| 2. Audiences | They’re still the people you write for most this month | Add or remove one |
| 3. Preferences | You’d still ask for each one if you were starting fresh | Delete the ones you’ve stopped caring about |
| 4. Project files | Only current versions, nothing superseded | Remove old drafts and old price lists |
| 5. Memory | Nothing sensitive and nothing out of date | Delete the entry |
| 6. Privacy | Nothing fails the Privacy Rule | Replace the detail with a description |
| 7. Rework | Recent answers needed less correcting than before | Find the line that’s steering it wrong |
Scoring: all seven pass, leave it alone. One or two fail, fix them today. Three or more fail, rewrite the block from the template above.
You can also get the model to do the first pass of the audit, as long as you make the decisions. This prompt asks it to find problems without rewriting anything yet:
| Here is the context block I use with you. Don’t rewrite it yet. First, list anything that looks vague, out of date, contradictory, or more detailed than an ordinary request would need. Second, list anything that looks like personal or confidential detail I should remove or describe more generally. Third, ask me up to three questions about my work whose answers would make this block more useful. [paste your block] |
Paste your block where the brackets are. You decide what changes; the model only points.
If you’re starting from nothing, this is the order I’d go in this week:
You can start all of this on your own. Where a whole team wants shared context blocks, building them together with someone who has seen how they go wrong shortens the learning curve, and that’s the kind of work our corporate workshops are built around.
Start with the request you make most often. Write down who it’s for and what they need from it, put that above the ask, and see what changes.
They overlap. I’d treat the prompt as the instruction for this request and context as the situation it sits in: who, for whom, why and within what limits. The practical difference is reuse: the instruction changes every time, while most of the situation stays put.
Long enough to change the answer and short enough to read in a minute. Past a page, it’s usually holding detail that belongs in a project file, or nowhere. The keep-or-cut test will tell you which.
It’s handy for small, stable preferences, and a weak place for anything you need to control, because you don’t always see what it’s drawing on. One detail from OpenAI’s Memory FAQ is worth knowing: deleting a chat doesn’t necessarily delete a saved memory created from it, so review the memory list itself.
Yes, if you keep it as plain text in a document you control and paste or attach it where you need it. Each tool stores context differently and those features change, so the copy in your own file is the one that survives a tool switch or a product change.
A context block is still useful, as long as it holds only non-sensitive detail: your role, your audiences described in general terms and your style preferences. Keep client, people and financial specifics out until there’s an approved tool and a policy that covers them.
I read these pages on October 21, 2026: Anthropic’s Prompting best practices guide and its Effective context engineering for AI agents post, OpenAI’s help pages on Projects, Memory and how data is used to improve models, the Claude Help Center page on creating and managing projects, and Google’s Use Gems in Gemini Apps page. I also read the abstract of the 2023 Lost in the Middle paper and use it only for the direction of its finding. I left out context-window sizes, prices and plan-by-plan limits, which change often. I did not test these features hands-on for this article. The seven kinds of context, the keep-or-cut test and the monthly audit are Future Factors working methods, not research-validated models, and the scenes are illustrations, not client stories.