6. Making It Stop Asking Who You Are

Every new conversation starts from nothing. You re-explain your job, your preferences and your context, or you accept generic output. Most people who drift away from AI do so here, not because the answers were bad but because getting good ones cost ninety seconds of typing every time.

There are five separate mechanisms for fixing that, they are constantly confused with each other, and they operate at different scopes.

The five layers

Layer	Scope	You control it
Custom Instructions	Every conversation	Directly, you write it
Memory	Every conversation	Indirectly, it decides what to keep
Projects	One workspace	Directly
Custom GPTs	A standalone assistant	Directly
Tone selector	One message	Directly, per message

The distinction that matters most is the first two. Custom Instructions are what you tell it about yourself, deliberately, once. Memory is what it decides to remember about you from conversations. One is a statement you wrote. The other is an inference it made, and you did not approve each one.

Start with Custom Instructions

Settings, then Customize ChatGPT. About 150 words, concrete, in two blocks. Vague instructions do nothing because it already believes it is being helpful and professional.

Who I am:
I run a 6 person design studio. Clients are mostly
US small businesses. I am not technical.

How to write:
Plain and direct. US spelling. No exclamation marks.
Never open with "I hope this finds you well."
Emails under 150 words unless I say otherwise.

How to behave:
If a question depends on facts you do not have, ask
me rather than assuming.
Flag anything in your answer I should verify.

Those last two lines are worth more than everything above them. They change the default from confident guessing to asking, which is the largest single quality improvement available from a settings box, and it costs two sentences.

Memory, and why to audit it

Memory saves things across conversations without you asking. That is convenient and it is also the layer to keep an eye on, for a reason people do not anticipate.

It records what it inferred, not what you said. Mention once that you are drafting something for a difficult client and it may keep a note that you have a difficult client. That note then colours unrelated answers weeks later, and you will not connect the two.

So audit it occasionally:

Settings, Personalization, Memory, Manage

You can read every entry and delete individually. Two reasons to actually do it: stale facts keep shaping answers long after they stopped being true, and anything wrong in there is wrong in every future conversation rather than one.

If you would rather it kept nothing, memory can be turned off entirely and Custom Instructions still work. That combination, explicit instructions on and inference off, is the setup we would pick for anyone who wants predictable behaviour.

Projects, for the work you repeat

A project holds a set of conversations, reference files and its own instructions. Worth setting up for anything you do more than twice a month.

What earns a place inside one:

• The documents you would otherwise re-upload. Given the three per day free limit from lesson 5, this alone can be the reason.
• Three examples of your own output. Showing beats describing, and it is the fastest way to stop output sounding generic.
• Your vocabulary. What "the report" means, your product names, which acronyms are yours.
• What not to do. "Never quote a price, tell me to check the rate card."

What does not: anything out of date. A project quietly feeding it last year's prices is worse than no project, because you will stop checking.

When instructions seem to be ignored

This happens, it is usually not a bug, and there are three causes in order of likelihood.

The rule was an adjective. "Be concise" is a mood, not an instruction. If you cannot check compliance by looking at the output, it will not hold.

Your message contradicted it. What you typed just now beats what you saved once, which is correct behaviour. Ask for a detailed breakdown and you will get one, word limit or not.

The block is too long. A wall of forty rules dilutes itself. When one rule keeps slipping, the fix is usually deleting four others rather than adding emphasis to that one.

The layer to reach for, by problem

Problem	Layer
It writes in the wrong register every time	Custom Instructions
It does not know what my company does	Custom Instructions
I keep re-uploading the same document	Projects
It keeps bringing up something stale	Memory, delete the entry
Just this one message needs to be blunt	Say so in the message

The bottom row is the one people over-engineer. A one-off does not need a setting, it needs a sentence.

Exporting and starting over

Everything above is recoverable and removable:

Settings, Data Controls, Export Data

That produces an archive including your memory entries. Worth doing once, partly as a backup and partly because reading what it has accumulated about you is more informative than any description of the feature.

Before the next lesson

Write your Custom Instructions now. Fifteen minutes, two blocks, and use the two behaviour lines verbatim if nothing better comes to mind. Then open Memory and read what is in there.

Next, the last lesson: which model to pick from a menu that changes every few months, and how to tell whether an answer came from a live search or from recall.