ChatGPT gives you four places to put standing information: custom instructions, projects, project files and memory. Most people use one of the four, usually the wrong one, and then wonder why it keeps forgetting things or bringing up the wrong context.
Here is what each one is for, with custom instructions examples you can paste, and the rule that decides where anything goes.
The four places, and what each is for
| Applies to | Put here | |
|---|---|---|
| Custom instructions | Every chat you have | Who you are and how you want to be talked to |
| Project instructions | Every chat in one project | The rules for one body of work |
| Project files | Opened when relevant | Reference material, style guides, past work |
| Memory | Carried across chats automatically | Things it picks up as you talk, which you edit rather than write |
The rule that resolves almost every case: scope decides location. If it is true everywhere, custom instructions. If it is true for one stream of work, project instructions. If it is a document, a file. If you would not want it applied to your next twenty messages, it is not standing information at all and belongs in the message you are typing.
ChatGPT custom instructions that work
The box asks two questions: what should it know about you, and how should it respond. Most people write a job title in the first and "be concise" in the second, which is why most people see no difference.
What actually changes the output is specificity about failure. Here is a version worth pasting:
What should ChatGPT know about you
I work in [role] at a [type of company]. Most of what I produce is
read by [audience], who are busy and not specialists in my area.
I already know [the basics of my field]. Do not explain those back
to me.
I am usually working from incomplete information and would rather
know that than get a confident answer built on a guess.How should ChatGPT respond
Answer first, reasoning after. If the answer is "it depends", say
what it depends on in one line before anything else.
No preamble. Do not restate my question back to me. Do not tell me
my question is a good one.
When I am wrong about something, say so directly in the first line.
When you are unsure, say how unsure and what would resolve it.
Never: bullet lists longer than five items, headings on anything
under 300 words, or a closing paragraph that summarises what you
just said. If I ask for a draft, give me the draft, not a draft plus
an explanation of your choices.The never list is doing most of the work there. Positive instructions like "be concise" compete with the model's defaults; explicit prohibitions replace them.
How to use ChatGPT projects
A project is a folder with its own instructions and its own files, and chats you start inside it inherit both. Create one, put the instructions in, upload the reference material, and every conversation in there starts with the context already loaded.
What projects are genuinely good for:
- One per client or account. Their files, their tone, their constraints, none of it bleeding into anyone else's work.
- One per recurring output. The weekly report, the monthly board update. Same brief every time, so the output stops drifting.
- One per long-running piece of thinking. A decision you keep returning to, a book you are writing, a job hunt.
What they are not for: a project per conversation. That is a chat with extra steps and a worse interface.
What belongs in project files
Reference material, not instructions. A style guide, three good examples of past work, a price list, a CV. The distinction is that instructions tell it how to behave and files tell it what is true.
Keep the file set small and current. A large pile makes retrieval less precise, and a stale document gets quoted back at you with total confidence because nothing in it says it is out of date.
ChatGPT projects vs custom GPTs
A custom GPT is built to be shared and reused, potentially by other people. A project is your workspace. If you are the only user, a project is simpler, holds files more comfortably and is easier to edit. Build a custom GPT when you want to hand something to a team or publish it.
There is a related question people ask about Claude projects, which work on the same principle with a heavier emphasis on the instructions block. If you use both tools, write the instructions once and paste them into each; it is plain text and nothing about it is proprietary.
What to do about memory
Memory picks things up as you talk and carries them between chats. It is useful and it drifts, because it stores what you happened to mention rather than what you decided mattered.
Two habits keep it useful. Read what it has stored every few weeks and delete what is stale or wrong, because a wrong memory is worse than no memory. And when something genuinely matters, put it in custom instructions or a project rather than hoping memory caught it.
Questions people ask
Why did my ChatGPT projects disappear?
Usually one of three things: you are signed into a different account, you are on a plan tier where the feature is not available, or the project was archived rather than deleted. Check the account first, since it is the most common and the easiest to miss.
Do project instructions override custom instructions?
They stack. Where they conflict, the more specific one generally wins, which is a reason to keep custom instructions about how you like being talked to and leave the work rules to projects.
How long should custom instructions be?
Both boxes together, under 400 words. Past that the later lines start losing to the earlier ones, and you will not be able to tell which part stopped working.
Can other people see my project files?
Not unless you share the project. A custom GPT you publish is different, and anything in its knowledge files goes with it, so check what is attached before publishing.
What is the difference between a project and a folder of chats?
The instructions and the files. Without those a project is only organisation; with them it changes what every conversation inside it produces.
Where to take this next
Paste the two custom instruction blocks above, edited for your situation, and give it a week. That single change does more than any prompt technique, because it applies to everything you type rather than to one message.
Then build one project for the thing you do most often. One, not six. Most people set up eight projects and use two.
If you would rather have the whole setup written for your actual job, that is what The Pack is: nine files, instructions plus eight task files, with install paths for ChatGPT Projects, Claude, Gemini Gems, Copilot and Grok.
Laisser un commentaire :