The Chat Settings Almost Nobody Opens, and What Each One Changes

Most people use a chat window like a search box: type, read, type again. The settings that change the quality of what comes back sit one or two clicks away, and most accounts never touch them. This is a tour of the ones worth knowing, starting with the one that costs nothing and pays back every day.

Standing instructions, the setting with the best return

Every one of these tools has a field somewhere in the settings for instructions that apply to every new conversation. Most people leave it empty and then retype the same preferences into individual chats for a year.

What belongs in it is not clever prompting. It is the boring and stable stuff: what you do, what you usually need, how long an answer should be, what to leave out. A single line telling it not to restate your question before answering saves you a paragraph in every reply you will ever get.

Editing your earlier message, and when not to

When an answer misses, the reflex is to reply with a correction. That works, but the poor answer stays in the conversation and everything that follows is written next to it.

The alternative is to go back and edit your own message, so the conversation continues from the corrected question instead. Claude and Gemini both offer this. In ChatGPT the option has not been consistently available to every account, and in the consumer version of Copilot it is generally not there at all, so check your own window before building a habit on it. Where it is missing, the same effect comes from starting a new chat and restating the question.

One clarification, because it is easy to overstate: editing removes the bad answer from what the model reads in this branch. It does not delete it. The earlier version stays in your account, and where a tool keeps memory across conversations it can still surface later.

There are also cases where following up beats editing, and they are common enough to matter. If the answer was half useful, editing throws away the good half with the bad. If the model ran a search, executed code, or worked on a file you uploaded, that result disappears with the branch and may not come back the same way. And if you are working through a problem step by step, the trail of steps is often the actual output.

When a fresh chat is the answer

Long conversations get worse, and the symptom is recognisable: the model keeps reintroducing something you ruled out twenty messages ago, or the answers drift towards the general and the vague.

The useful mental model is that the conversation is a budget, not an archive. Everything still in play competes for the model’s attention, and once a thread outgrows what the model can hold, older parts get compressed or dropped rather than read in full. Either way, your false starts are being paid for. When you notice the symptom, the fix is not a sharper instruction. It is a new chat containing the one paragraph that mattered.

The search toggle, and a trap inside it

With web access off you are asking what the model absorbed during training. With it on, you are asking it to go and look. That difference matters for anything with a current state: prices, versions, who holds a job, whether a product still exists.

The obvious tell is that an answer with no links probably came from memory. The less obvious one, and the more dangerous, is that links are not proof of the opposite. A model working from memory can produce a citation that looks entirely ordinary and leads nowhere. If a link matters, open it.

Temporary chats: useful, but not what the name suggests

Most of these tools now offer a mode where the conversation stays out of your history. It is worth knowing where that switch is before you need it. It is also worth knowing what it does not mean.

OpenAI states that a temporary chat will not appear in your history and will not be used to improve its models while it stays temporary, and adds that for safety purposes a copy may be kept for up to 30 days. Google keeps temporary Gemini chats with your account for 72 hours, does not use them to train its models, and says plainly that even with temporary chats it still uses your conversations to help protect Google, its users and the public, including with help from human reviewers. Where a conversation has been through human review, Google retains it for up to three years and deleting your activity does not remove it.

So the honest summary is: temporary means out of your history and out of training, not gone. That is a real benefit and worth using. It is not a reason to paste in something you would not otherwise hand over. If a document, an email or a group chat screenshot belongs to someone else, the right move is to remove the names and identifying details first, or not to upload it at all.

Memory, and what the list does not show you

Several tools now carry facts about you from one conversation to the next. Opening that list once is worth the two minutes, because there are usually entries you did not intend to leave behind.

Do not mistake the list for the whole picture, though. OpenAI says outright that its summary of what ChatGPT remembers will not include everything it draws on from your past chats, and Gemini’s broader personal context is a switch rather than a readable list. So the list is where you can remove specific things, not a complete record of what the model knows about you.