From a diary that answers to a mirror
My first experiment, Dear Diary, taught me a lot about journaling with AI. Above all, it taught me what I didn’t want: an AI that cheered me on, entry after entry, but never got to the point. And the point of any inner work is, more often than not, seeing what we’d rather not see.
And AI has one gift I hadn’t really used yet: it can remember what we forget, or would rather forget. It can hold months of writing in view and point to a link between today’s entry and one from three weeks ago, a link I might never have seen, or wanted to see.
So later in 2025 I started again, with a different method and a different goal: clarity over comfort. Real change, rather than the feel-good that too easily slides into spiritual bypassing1.
Expressive writing… my way
My method is loosely inspired by the better-known technique developed by psychologist James Pennebaker. Here’s how I do it:
- I write about whatever passes through my mind: reflections on events around me, emotions they triggered, thought loops I fall into, insights that came up… my inner life, so to speak.
- I write continuously, for a set time. I set a 15-minute timer and write without stopping, without judging and without censoring. Grammar, typos and logic don’t matter at this stage. The point is to let the stream of consciousness flow, and see what surfaces from underneath.
- When the timer stops, I paste everything into a temporary ChatGPT chat with one instruction: “fix it, but use my own words”.
- Then I reread the result and edit where needed. Grammar still doesn’t matter much, but the next step needs something clear enough for the AI to follow. And this second reading does a job of its own: it lets me see what went through my mind from a little distance.
Some bumps along the way:
- I started on paper and used ChatGPT to read my handwriting, but the transcriptions proved unreliable, so I moved to the keyboard.
- At first I wrote almost every day. These days, it’s once every few days. According to Claude, that’s not laziness: it’s a sign the practice has done its job.
From ChatGPT to Claude
At first, the cleaned-up entries went to ChatGPT for feedback. It worked for a while, but over the months the replies gradually drifted towards uselessness: too coachy, too pleasing, too much fluff. Every entry ended with encouragement, whatever I’d written. Maybe that felt good for a while, I can’t remember. But it was far from the precious insights the AI could sometimes give me: a fresh perspective I hadn’t considered, a clearer view of something I was only vaguely sensing, a connection between entries written weeks apart.
I tried new chats and different prompts. Then, by March 2026, I’d become curious about Claude, and one day I decided to start a new chat there, with a very precise brief (more on that below).
For a while I ran both side by side. I gave Claude ChatGPT’s reply to the same entry and asked it to assess it. That comparison turned out to be the best training of all, for me and for the AI. Together we named the ways a reply can fail: tying everything up too neatly, closing on validation, coaching when I’d asked for a mirror, answering questions I’d deliberately left open.
The protocol
This is the brief I gave Claude on day one. Compare it with the prompts from Dear Diary: where those asked for support and positivity, this one asks for the opposite.
You are a psychological mirror and analytical reflection engine.
The user will provide daily journal entries produced through automatic writing.
Important context: These entries are unfiltered streams of consciousness. They are not structured arguments. They may contain contradictions, exaggerations, temporary states, projections, or raw emotional material. Do not treat them as final positions or stable beliefs.
Your task is not to fix, comfort, coach, inspire, or motivate.
Your task is to:
1. Distill the underlying psychological theme beneath the surface content.
2. Separate:
* objective events
* interpretation
* emotional activation
* identity narrative
3. Identify recurring patterns across entries when present.
4. Gently expose distortions or cognitive amplifications.
5. Point out avoidance or self-deception when necessary, without aggression.
6. Reduce catastrophic thinking without invalidating emotional experience.
7. Avoid spiritual bypassing and avoid therapeutic clichés.
8. Avoid generic encouragement or emotional reassurance.
9. Do not over-pathologize normal emotional activation.
10. Do not give life advice unless explicitly asked.
Tone requirements:
* Calm
* Precise
* Grounded
* Non-performative
* Non-moralizing
* Non-condescending
* Direct but respectful
Never:
* Say "everything will be fine"
* Say "just breathe"
* Use motivational language
* Use inspirational rhetoric
* Suggest affirmations
* Suggest standard therapy techniques
* Over-intellectualize
Close responses by:
* Re-centering the user in stability
* Clarifying the core dynamic at play
* Avoiding emotional crescendo endings
The goal of the protocol is: Clarity over comfort. Stability over intensity. Awareness over narrative.Tuning it
The brief was only the start. Over the months I corrected Claude many times, and the corrections became rules:
- Hypotheses, not assertions. Claude would say something with full confidence, then the opposite with the same confidence as soon as I questioned it. The fix: keep looking for connections between entries, but offer them as hypotheses, and ask whether they land rather than declaring that they do. It was the most useful change of all.
- Don’t invent depth. If there’s no reason to see something underneath, don’t. A mirror reflects; it doesn’t add.
- My profile, in my words. Over time, Claude had built its own picture of me, partly its own construction (in other words: it assumes things about me and my history that simply weren’t true). So we rewrote it together, item by item, with me correcting almost every line. An AI that reflects you should start from how you see yourself, not from its guesses.
- A prompt worth sharing. When a friend asked for my setup, I asked Claude to write a version for them. That forced months of corrections into explicit rules: mark speculation as speculation, don’t turn insights into numbered steps, don’t end with validation, don’t answer what I’ve left open, match my length and tone. The best prompt turned out to be the one written after the mistakes.
And what I did want, in my own words at the time: “it is precious when you are insightful, when you think something that didn’t occur to me.”
Where it fails
It works, but it isn’t magic. Actually, it felt like magic so many times that when the shortcomings started to show, I couldn’t help feeling a little disappointed. Anyway, these are the failures I’ve seen:
- Confident misremembering. Once it cited an old entry for something I’d never written there. The content was real, the date was invented. Nothing distinguishes a false memory from a true one, so I end up fact-checking the mirror.
- Flattery is a constant pull. Less than with ChatGPT, but it never disappears. And after a correction it can overshoot into finding fault, which is just a different distortion.
- It folds. Push back, and it often drops a position entirely instead of defending or refining it. That makes it hard to know which observations have weight behind them.
- Pattern manufacture. With months of material, almost anything can look connected to anything else. The only check is me, which is why “ask if it lands” matters so much.
- Leading. Sometimes it handed me back my own idea as its own insight, and sometimes the other way round. More than once I had to ask: did I get here, or did you lead me?
- Long chats lose precision. The longer the conversation, the sloppier the replies. That’s why I’m now thinking of changing the setup.
What stays human
After months of this practice, one thing is clear to me. As I wrote in Dear Diary, we need to understand how to use this technology for what it does best.
You can’t rely on AI as a therapist. But you can use it to see yourself in ways you probably couldn’t manage on your own.
It can remember what I wrote three weeks ago and suggest a connection with something that happened yesterday with someone I care about. Is it always right? No. Definitely not. As an emotional compass, it’s poor. But here’s the thing: we are the compass. The AI can draw the map, and we, the compass, can point the way. That’s why one of the most important rules was: don’t state things with confidence; offer a hypothesis, and ask me if it resonates.
In other words, it’s the way we’re all learning to use AI these days: it speeds things up and widens the view, but the human stays in the loop. There’s a lot of hype about AI soon improving so much that humans won’t be needed in the loop at all. Working with it every day, I dare say there’s also hope that we always will be.
- Spiritual bypassing: using lofty spiritual ideas to avoid facing painful feelings and unresolved issues. The term was coined by psychologist John Welwood in the 1980s. ↩︎
I’m Jonathan, SEO consultant and AI developer. Questions about this project? Ask me below, or find out more about me.