How to Give ChatGPT Context (Without Repeating Yourself)

TL;DR: ChatGPT has six places that will hold context for you: the prompt, custom instructions, project instructions, project and chat files, memory, and Custom GPTs. Rank them by durability, not by novelty. Custom instructions cap at 1,500 characters on Free and Go and 5,000 on Plus and Pro. Project instructions override global ones. Memory is lossy.

If you have pasted the same three paragraphs about your company, your codebase or your writing voice into a fresh chat for the fourth time this week, you know the problem. Learning how to give ChatGPT context is easy. Learning how to give it once is the part nobody explains, because the answer is not one feature. It is six of them, they behave differently, and OpenAI documents them in five separate places.

This guide ranks each mechanism by durability, and says what it is good at, what it is bad at, and where it fails without telling you. Every limit here was read from OpenAI's own documentation on 28 August 2026, with the URL. Where OpenAI publishes no limit, this page says "not published" rather than guessing.

What are the six places ChatGPT will hold context?

Six, and they are not interchangeable. Durability is the useful ranking: how long context survives without you touching it again.

The ordering is deliberate. The most durable mechanism is the least clever: a block of text you own, stored outside ChatGPT, that you can paste into anything. Everything below it is a convenience, and every convenience has a plan gate, a cap or a failure mode.

Does putting context directly in the prompt still work?

Yes, and it is the highest-fidelity option. Nothing summarises it, nothing overrides it. The cost is that it dies with the conversation.

OpenAI's consumer prompting guide names four parts worth including for anything larger than a one-liner: "Goal", "Context", "Output" and "Boundaries" (learn.chatgpt.com/docs/prompting.md, accessed 28 August 2026). The developer guide is specific about placement: context "is usually best positioned near the end of your prompt, as you may include different context for different generation requests" (developers.openai.com/api/docs/guides/prompt-engineering.md, accessed 28 August 2026).

That second point is the one people get backwards. Stable instructions go near the top. The variable material — this week's numbers, this client's brief — goes near the bottom.

Block 1. The general shape

GOAL
<what you want produced, in one sentence>

CONTEXT
<who you are, what the project is, what already exists>

OUTPUT
<format, length, level of detail, who reads it>

BOUNDARIES
<what must not change, what to flag instead of guessing>

Block 2. A filled-in version

GOAL
Draft the launch email for our September pricing change.

CONTEXT
B2B SaaS, 2,100 paying customers, mostly solo founders and small agencies.
Voice: direct, no hype, short sentences. We have never raised prices before.
Existing customers keep their current rate for 12 months.

OUTPUT
One email, 180-220 words, plain text, subject line plus three variants.

BOUNDARIES
Do not invent a discount or a deadline. Flag anything you are unsure about
instead of guessing. Do not use the word "excited".

Block 3. The "restate before you answer" check

Before you answer, restate in three bullets: who I am, what I am asking for,
and any constraint you think applies. If any of the three is a guess, say so
and ask me one question. Then wait for my confirmation.

Block 3 is worth a habit. It catches the case where ChatGPT has quietly picked up stale context from elsewhere and is about to write 800 words on the wrong premise.

Where it silently fails: long conversations. Everything you paste competes for the same context window, and past a certain length, early material stops carrying the weight it did at the start. The fix is usually a new chat with the block re-pasted, not another correction inside the old one. The symptoms are covered in why ChatGPT repeats itself.

What can custom instructions actually hold?

Custom instructions are the account-wide layer. They apply to every new chat without you doing anything, which makes them the highest-leverage place for anything true all the time.

The limits are published, which is unusual, and worth quoting exactly. From OpenAI's custom instructions article (help.openai.com/en/articles/8096356-chatgpt-custom-instructions, accessed 28 August 2026): "Free and Go users can save up to 1,500 characters in custom instructions." And: "Plus, Pro, Enterprise, Business, and Education users can save up to 5,000 characters." The same page states that "Custom instructions are available on all plans on Web, Desktop, iOS, and Android."

1,500 characters is roughly 250 words, which is smaller than most people assume.

Block 4. Standing facts about you, generalist

I run a two-person B2B SaaS business. Non-native English speaker, but I write
in English professionally, so do not simplify vocabulary.
I care about: accuracy over encouragement, short sentences, concrete examples.
When I ask about a product, tool or price, I want the source and the date.

Block 5. Response rules, generalist

Answer first, explain after. No preamble, no summary of my question.
No bullet lists unless I ask for them or the content is genuinely a list.
If you are not confident, say which part you are unsure about.
Never invent a statistic, a citation, a price or a quote. Say "not published"
instead of estimating.
Do not compliment the question.

Block 6. Developer variant

Stack: TypeScript, Next.js App Router, Postgres, Tailwind. Package manager: npm.
Show diffs or full files, never fragments with "..." in the middle.
Do not refactor code I did not ask you to touch.
If a library API changed recently, say so instead of using the version you know.
Comments only where the reason is non-obvious.

Block 7. Writer and marketer variant

I write for founders and operators, not for a general audience.
Banned: "delve", "leverage" as a verb, "in today's fast-paced world", "unlock",
"game-changer", em-dash-heavy sentences, and rhetorical questions as openers.
Prefer specifics to adjectives. One idea per sentence.
British spelling. No exclamation marks.

Block 8. Researcher variant

Field: applied linguistics. I read primary sources, so give me the paper, the
year and the authors, not a summary of the consensus.
If a claim comes from a single study, say so. If you cannot locate a source,
say you cannot rather than describing what is probably true.
Distinguish "the paper found" from "people often say".

The practical consequence is ours, not OpenAI's. If you edit your custom instructions and the change matters, start a new chat. OpenAI does not say to do this and nothing on their page recommends it. It is simply the habit that stays correct under either version of their own documentation, which is the only reason to adopt it.

Where it silently fails: three ways. The cap is easy to exceed, because 1,500 characters is about a quarter of a page. Custom instructions are account-scoped, so they do not travel to Claude, Gemini or an API call. And they never appear in the transcript, so when an answer comes out oddly shaped, the cause is somewhere you are not looking. How to make ChatGPT follow your instructions goes deeper on that diagnosis.

When should you use a Project instead of custom instructions?

When the context is true for one body of work and false everywhere else. A project scopes instructions, files and chats to a single effort.

Two facts from OpenAI's Projects article (help.openai.com/en/articles/10169521-projects-in-chatgpt, accessed 28 August 2026) change how you should use them. First: "Projects are available to all free and paid subscription types globally", and "Users can create an unlimited amount of projects." Second, and much more important:

"Project instructions only apply inside the respective project and will override your global custom instructions"

Read that twice. A project is not additive. Inside it, the project instructions win. If your global block contains something you still need, such as your banned-words list or your "no preamble" rule, you have to restate it in the project.

That one sentence explains a common complaint: someone sets up a project and ChatGPT stops obeying tone rules they set months ago. Nothing broke. The project replaced them.

Block 9. Project instructions, client work

This project is for the client ACME (mid-market logistics, UK).
Their voice: plain, slightly formal, no Americanisms, no exclamation marks.
Their audience: operations managers, not executives.
Always use "colleagues", never "employees". Always spell it "programme".
Deliverables are drafts for internal review, never final copy.

Carry over from my global instructions: answer first, no preamble, never
invent a statistic, say "not published" rather than estimating.

Block 10. Project instructions, codebase

This project is the payments service. TypeScript, Node 22, Postgres 17,
Stripe API. Tests are Vitest. We do not use an ORM; queries are raw SQL.

Rules: never suggest a schema change without also writing the migration.
Never rename an exported symbol. Treat every column as load-bearing.
If you need to see a file, ask for it by path instead of assuming its contents.

Carry over from my global instructions: full files not fragments, no
unrequested refactors.

Block 11. Project instructions, recurring report

This project produces the Monday status update for the leadership team.
Sources: the plan doc and the metrics CSV in this project's files.
Format: one page. Decisions needed first, then progress, then risks, each with
an owner and a date.
Never change approved dates or budget figures. Flag conflicts instead of
resolving them. Do not add commentary that is not in the sources.

Block 12. Project instructions, research

This project is my literature review on <topic>.
Only use the PDFs uploaded to this project. If an answer is not in them, say
"not in the uploaded sources" and stop.
Every claim gets a file name and a page or section reference.
Never paraphrase a finding into a stronger claim than the paper makes.

Where it silently fails: the override, most of all. Beyond that, "Chats created with a GPT can’t be moved into a project", and project-only memory means "Your previously saved memories are not referenced in project chats". Both quotes come from the same article, and neither is announced in the interface at the moment it bites you.

Is ChatGPT memory a reliable place to keep context?

No. It is a convenience, and treating it as storage is the most common mistake in this area.

Memory is genuinely useful. OpenAI describes it as remembering "useful context from your chats, files, and connected apps to personalize your experience, so you don’t have to repeat yourself as often" (help.openai.com/en/articles/8590148-memory-faq, accessed 28 August 2026). Note the wording: "as often". Not "never".

Here is why it is not a filing cabinet, in OpenAI's own words on that page:

• The visible summary is not the whole thing. "While the memory summary should capture the most important details, it will not include everything that ChatGPT remembers based on your chats."
• It is a synthesis, not a list. "ChatGPT’s memory is based on a continually updated synthesis of context from your past chats, which may be broader than what can be shown as individual items in a summary."
• The sources panel is partial too. It "may not show every factor or source that shaped a response".
• Deleting memories does not delete their origins. "This does not delete your past chats." And: "If you turn memory back on later, ChatGPT may create new memories from chats that remain in your chat history, including older chats."

A summary of a conversation is not the conversation. That matters most in exactly the case people rely on memory for: you explained something in detail three weeks ago, and what survives is a compression that may have dropped the qualifier you cared about.

Block 13. Audit what it thinks it knows

List everything you currently believe about me and my work, grouped into:
(1) my job and company, (2) my projects, (3) my preferences, (4) anything else.
Mark each item as "certain" or "inferred". Do not soften anything.

Block 14. Correct a specific stale item

Update what you know: I no longer work on <old project>. That work ended in
<month>. Do not reference it again unless I bring it up. My current focus is
<new thing>.

Block 15. A clean-slate chat without touching settings

Ignore anything you remember about me for this conversation. Use only what I
write in this chat. If you find yourself about to reference something I have
not said here, ask me instead.

To inspect and clear memory, go to the settings, not the chat: "Memory controls are available in Settings > Personalization > Memory", and the memory summary is editable there. Two things to know before you rely on a purge. "Temporary Chats do not use existing memories or create new memories", which is the cleanest way to work without it at all. And a full removal is more work than it looks, because to delete something ChatGPT may know, per the same article, "you’ll need to delete every source where it appears, including past chats, archived chats, files, the memory summary, and disconnect any connected apps that may contain that information".

Where it silently fails: it never tells you it is wrong. A stale memory produces a fluent, confident answer on a premise you retired months ago, and nothing flags it. If ChatGPT seems to have forgotten something instead, why ChatGPT forgets what you told it covers the other direction.

Can you still build a Custom GPT for reusable context?

Probably not, and this invalidates the advice in most older articles.

From OpenAI's GPT documentation (help.openai.com/en/articles/8554397-creating-and-editing-gpts, accessed 28 August 2026): "New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro." The same page adds: "Existing GPTs remain available to use and can still be edited if existing plan and permission requirements are met." Creation and editing remain available in Business, Enterprise and Edu workspaces, subject to workspace settings.

So if you already have a Custom GPT, keep using it. If you were about to build one as your reusable-context solution on a Plus account, that route is closed, and any guide saying otherwise has not been checked recently.

A second fact matters even more if you do have one, from the GPTs article (help.openai.com/en/articles/8554407-gpts-in-chatgpt, accessed 28 August 2026): "GPTs do not use saved memory, custom instructions, or previous conversations. Each conversation starts fresh."

That is the cleanest isolation on this page, and it cuts both ways. A GPT is never polluted by stale memory. It also completely ignores the custom instructions you spent an afternoon writing.

Block 16. Custom GPT instructions, if you can still edit one

You are a <role> for <company>, a <one-line description>.

ALWAYS
- Answer in <format>. Lead with the recommendation.
- Use only the uploaded knowledge files for facts about our product.
- If the files do not cover something, say so instead of generalising.

NEVER
- Invent pricing, dates, customer names or metrics.
- Use marketing language. Plain sentences only.

WHEN <situation> HAPPENS
- Do <specific step>, then ask me before continuing.

When do uploaded files beat pasted context?

When the material is longer than you want to read, and when several chats need the same source of truth. Files attach to a project or a single chat, and unlike a pasted block they cost no typing.

The published file counts, from the Projects article (28 August 2026), are "Free: 5 files per project", "Go, Plus: 25 files per project" and "Edu, Pro, Business, Enterprise: 40 files per project". A GPT is separate: "You can attach up to 20 files to a GPT. Each file can be up to 512 MB."

Where it silently fails: visually. Per the File Uploads FAQ (28 August 2026), outside ChatGPT Enterprise, "All other plans and document files only support text-based retrieval. This means that ChatGPT will extract digital text from the file and discard any images." So a brand guideline PDF that expresses everything through swatches and layout arrives as almost nothing, and a scanned contract as nothing at all. Neither errors. You get a confident answer built on whatever happened to be selectable text.

The fix is not a better file. It is a short written summary beside it.

Block 17. The file companion note

Attached: <filename>. It is a <type of document>.
What matters in it: <2-4 lines of the actual substance>.
What to ignore: <appendices, boilerplate, anything visual>.
If you cannot find something I have described here, say so — do not infer it.

Block 18. A source index for a project

Sources in this project and what each is for:
- brand-voice.md — tone rules. Authoritative on style.
- pricing-2026.csv — current prices. Authoritative on numbers.
- old-deck.pdf — background only. Superseded, do not quote figures from it.
Rule: when two sources conflict, the one marked authoritative wins, and tell
me the conflict happened.

How do you build a reusable context block you can paste anywhere?

Write it once, keep it somewhere you can reach in two seconds, and paste it at the top of any chat that needs it. This is the most durable mechanism on this page and the least impressive-sounding.

It is also the only portable one. A pasted block works in ChatGPT, Claude, Gemini, an API call and a tool that does not exist yet. Custom instructions, projects, memory and GPTs are features of one product.

A plain text file works, and it costs nothing. Keep context.txt on your desktop or a pinned note on your phone. That version has no sync, no search and no way to update ten prompts at once, but it beats retyping, which is the whole point.

Block 19. The master context block

=== STANDING CONTEXT — updated <date> ===

WHO
<name>, <role> at <company>. <company> is <one line>. Team size <n>.

WHAT WE SELL
<product>, to <audience>, priced <how>. Current focus: <thing>.

VOICE
<3-5 concrete rules, including banned words>

WHAT YOU SHOULD ASSUME
<the 3 facts you re-explain most often>

WHAT YOU SHOULD NEVER DO
<the 3 mistakes you correct most often>

=== END STANDING CONTEXT ===

Block 20. Company and brand slice

COMPANY: <name>, founded <year>, <n> people, bootstrapped/funded.
CATEGORY: <what a stranger would call us>
POSITIONING: we are the <x> for <y> who <z>.
NOT: <the two things people wrongly assume we are>
PROOF WE CAN USE: <numbers you have actually verified, with dates>

Block 21. Product slice

PRODUCT: <name>. Does <one sentence>.
PLANS: <names and what separates them>
SHIPPED: <features that exist>
NOT SHIPPED: <features people ask about that do not exist — never imply these>
KNOWN WEAKNESSES: <the honest list; use these when asked for drawbacks>

Block 22. Audience slice

PRIMARY AUDIENCE: <who>, <seniority>, <what they are measured on>.
THEY ALREADY KNOW: <do not explain these>
THEY DO NOT KNOW: <explain these every time>
WHAT MAKES THEM CLOSE THE TAB: <2-3 things>
WORDS THEY USE FOR THIS PROBLEM: <their language, not yours>

Block 23. Codebase slice

REPO: <name>. <framework> <version>, <language>, <database>.
CONVENTIONS: <where things live, how things are named>
DO NOT TOUCH: <migrations, generated files, vendor code>
TESTS: <runner, how to run them>
DEPLOY: <what happens on merge — say if it is straight to production>

Block 24. Style guide slice

SENTENCES: short. One idea each. Active voice.
BANNED: <your list>
FORMATTING: <headings, lists, tables — when each is allowed>
NUMBERS: always with a source and a date. Never estimate.
TONE CHECK: would a sceptical reader in a hurry finish this?

Block 25. The two-line version for quick chats

Context: I am <role> at <company>, a <one line>. Audience is <who>.
Voice: <three words>. Never <the one thing>.

Block 26. The handoff, for when a chat has gone long

Write a handoff block I can paste into a fresh chat. Include: the goal, the
decisions we made and why, the constraints, and what is still open. Facts only,
no narrative. Under 200 words.

Block 26 is how you move context between conversations without losing the reasoning, and it pairs with knowing when to start a new chat versus keep going.

Once the block exists, the only question is where it lives. A note is free and works. A tool adds three things a note cannot: update one value and have every future prompt pick it up, keep the block reachable inside the ChatGPT page itself, and share it with other people.

That is the job Prompt Architects does. Our /faq page describes global variables as "reusable values you define once and inject into any prompt using {{variable_name}} syntax", and a saved context is attached to a prompt "with one click and it's injected at enhancement time". Being straight about the tiers, checked on 28 August 2026: the Free plan "includes 5 prompt enhancements per day, forever. No credit card required", but the Personal Context Library is paid, appearing on Pro, Advanced and Team in the /pricing comparison table, and shared team context is a Team feature. If you need one block for yourself and you are happy pasting it, the text file is genuinely fine, and we would rather say so.

One more surface, worth knowing about and worth not planning around: Skills. OpenAI describes them as reusable workflows that can include instructions and files, but states that "Skills are available to eligible ChatGPT Business, Enterprise, Healthcare, and Edu users, subject to workspace settings and product availability" (help.openai.com/en/articles/20001066-skills-in-chatgpt, accessed 28 August 2026). Not an option on a personal account today.

Which mechanism should you actually pick?

Work down this list and stop at the first line that matches.

1. True in every chat, forever? Custom instructions. Keep it under 1,500 characters if you are on Free or Go.
2. True for one body of work? A project, with project instructions — and restate anything from your global block that still matters, because the project overrides it.
3. Long reference material several chats need? Project files, plus a written companion note saying what is in them.
4. True only today? The prompt, near the end of it.
5. Needs to work outside ChatGPT, or outside this account? A written context block you own.
6. Anything that has to be correct? The written block, again. Not memory.

Memory is deliberately not on that list as a destination. Leave it on, because it removes friction. Just do not put anything there you would be annoyed to lose or embarrassed to see resurface. For the deeper version of this discipline, advanced ChatGPT prompting picks up where this leaves off.

The habit underneath all six mechanisms is identical. Write your context down once, in a place you control, precisely enough that a stranger could act on it. Every feature here is a way of delivering that text. None will write it for you.