How to use ChatGPT for journaling (without it just agreeing with you)

Why venting to ChatGPT feels great for ten minutes and then feels like nothing

You type the whole mess out at 1am. It comes back warm, organised, and a little bit impressed with you. Then you close the tab and notice nothing actually moved. You do not understand yourself any better than you did twenty minutes ago. You just got told you were handling it well by something that was going to say that either way.

Cutting to the chase: ChatGPT is genuinely good at this, but not on default settings. Out of the box it is tuned to be agreeable and to jump straight to solutions, and those are the exact two behaviours that wreck a journaling session. Everything below is how to turn most of that off. It takes about four minutes and you only do it once.

The agreeableness is not in your head — it is in the training

In April 2025 OpenAI shipped a GPT-4o update and pulled it four days later because the model had become, in their own word, sycophantic. It was agreeing with more or less everything, including things nobody should be agreed with. Their postmortem is worth reading for one admission in particular: they had leaned too hard on short-term feedback. The thumbs-up. The "that was helpful" click.

They rolled that update back, but the structural thing did not go anywhere. A general-purpose assistant is optimised to make you feel like the answer was good. Most of the time that is completely fine — you wanted the email drafted and now it is drafted. When the subject is you, "made you feel good about the answer" and "actually helped you" stop being the same thing, and the model has no reliable way to tell them apart. The friend who tells you what you want to hear is not your most useful friend either. Same problem, more compute.

Fix it once: custom instructions

In ChatGPT this lives under Settings, then Personalization, then Custom instructions. There is a box asking how you want ChatGPT to respond. Paste something close to this into it, in your own words if you prefer:

When I am thinking through something personal: do not open with praise or reassurance, and do not tell me my feelings are valid. Ask me one question at a time and wait for my answer before asking the next. Do not offer advice, solutions, or a plan unless I ask for them directly. If I am being vague, say so and ask me to be more specific. At the end of the conversation, summarise what I actually said in three sentences, using my words, and do not add anything I did not say.

Every line in there is doing a job. "One question at a time" is the big one — without it you get a numbered list of six questions back, which is not a conversation, that is a form. "Do not tell me my feelings are valid" removes the single most hollow sentence in the entire genre. And the summary rule at the end stops it from quietly inventing an insight you never had and handing it back to you as if you said it, which it will absolutely do if nothing stops it.

The opening move: dump first, ask second

Most people start with "help me process this" followed by a tidy three-line version of the problem. That gets you a tidy three-line answer. The tidying is exactly where the useful material went missing.

Do it backwards. Dump the messy version first — unedited, out of order, petty parts left in, no attempt to make yourself sound reasonable. Then one instruction: "Do not respond to that yet. Ask me one question about it." What you are after is the follow-up, not the reply. The reason talking to a good listener beats writing in a notebook has almost nothing to do with their advice. It is that a notebook never says "wait, say that part again."

If typing makes you edit yourself, use the microphone and talk instead. People say truer things out loud than they type. Typing has a backspace key and your inner PR department knows exactly how to use it.

Make the session end somewhere

James Pennebaker spent decades studying what makes writing about hard things actually useful. When his team ran the real words through text analysis, the thing that predicted improvement was not how much emotion people poured out. It was whether the writing changed shape across sessions — from fragmented venting into something that hung together, carrying more of the words people use when they are working something out. Because. Realise. The reason. Understand. The venting was the raw material. Making sense of it was the part that did the work.

So finish on purpose instead of drifting off. Two prompts do it: "What did I actually say in there — three sentences, my words" and "What is one question I should sit with tomorrow." That is the whole ritual. If you close the tab holding a paragraph you could not have written at the start, that was a real session. If you close it with a fresh replay of everything that went wrong plus a compliment, that was an hour of chewing. There is a post here on the difference between reflecting and ruminating that goes further into where that line sits, because it is easy to cross without noticing.

Three things this setup still will not fix

Continuity. ChatGPT has memory now and it is genuinely useful, but it is a summary of you rather than a record of you — it keeps what it judged worth keeping and drops the rest. Do not count on it to spot the pattern you have been circling for six weeks, because it may not be holding six weeks of anything. If you want that, keep your own running note somewhere and paste it back in at the start.

Privilege. Whatever you type into a general assistant lives inside a general-purpose company: its retention policies, its product decisions, its lawsuits. In 2025 a court order in the New York Times case forced OpenAI to preserve chats it would normally have deleted, including ones users had deleted themselves, and OpenAI’s own CEO said plainly that none of it carries the legal privilege you get with a doctor or a lawyer. At an absolute minimum, open Data Controls and switch off using your chats to improve the models, and never use the share link for anything personal. There is a longer post here walking through that whole episode if you want the detail.

A floor. If a session goes somewhere genuinely dark, a general assistant has no real plan beyond pasting a hotline number and carrying on. That is not a bug in your prompt. Nobody built it for that, and it is better to know that now than at 3am.

The honest bit, since I build one of these

Take this with the appropriate salt: I make a tool in this space, so I am not a neutral party. Haven exists because I wanted the setup above without configuring it every time — a room that already asks one question at a time, does not have a general-purpose company and its lawsuits sitting underneath it, and knows what to do when a conversation turns. It is a space to think out loud. It is not therapy, it does not diagnose anything, and it will not fix you. There is a page here on where it sits alongside professional support, and a blunter one listing what it flatly will not do, which is the more useful of the two.

And if ChatGPT plus four minutes in the settings menu is enough for you, use that. Seriously. It is free and it is already open in another tab.

Where none of this is the right tool

Everything above is for the ordinary weight of being a person: the bad day, the decision you keep circling, the conversation you keep rewriting in your head at the gym. It has limits. If the same dark loop has been running for weeks and it is into your sleep and your appetite, no prompt fixes that, and no amount of clever custom instructions makes a chat window into the right room. That is where a real therapist earns the hour. If I believe that, it is because of exactly one of them — I went through three who were not a fit before I found the one who was, and he did not just listen, he helped me work through it. If your first one was a nodding wall, that was fit, not proof the whole thing is useless.

And if any of this takes you somewhere unsafe — if you are thinking about hurting yourself — close the chat and reach an actual person. In the US you can call or text 988, the Suicide and Crisis Lifeline, any time. That is precisely what it is there for.