.png)
2 hours ago
1
Last year, among over things, me and my dear friend Gianlu built Deeply: an AI diary that wrote back to you. For a number of reasons we had to shut that down, but I still personally use AI to achieve similar outcomes and I’ve made progress on how to use it so that it’s more and more useful and surprising.
This is an AI use case that worked well for me and that friends ask for every time I mention it.
This assumes you have a collection of personal notes/journal that you consistently write. If you don’t, you can still try to get insights on new ways to play around, but you may not be able to reproduce results as good as mine. When you’re writing for no purpose at all, your brain reveals information that may not come out when you’re writing for a specific purpose and/or when you know you want to use what you’re writing as input for AI.
1. Export all your notes/journals in text (or markdown) format and group them in a single directory.
If you use Obsidian, you’re already good to go. I personally just have a git repo and write markdown in vim. I used to take all my notes on Notion but when Notion AI came out I didn’t like the output and I wanted to use other AI tools on my notes, so the only option was to export the notes away from the platform. This is how I ended my journey on Notion to go back at simple text files. Less is always more. Simple is timeless.
2. Open your notes directory with Cursor/Claude AI or your favourite AI tool that has cli/file-system access. If you’re extra-paranoid about privacy I’m sure you already figured out your workflows to do all this jazz in an extra-private and safer way.
3. To get started, populate the context with the structure of the dir and have the model figure out useful details about you.
Something like
Read this repo’s README, analyze its structure, read the core docs. Tell me about {{your name here}}. what’s their life like? what in this notes/journal repo?
will do.
Alternatively, add the relevant information in the context yourself, by telling the model where in the dir will it find what, basics about you, relevant files, etc.
Key docs I keep that I think are relavant for quality output here include:
README.md: I use the same dir for daily journal and long-living notes, so I talk about both the purpose and dir structure of it all in the README — if your collection of notes is simple enough, you won’t need this
timeline-of-my-life.md: a file that has 1 line per year of my life since I was born, with 1 key event of that year and general feelings/worries
4. Let’s add the actual journals content and their metadata (even the frequency of your journaling can tell something about your life in given periods!) to the model context.
Something like
List all their recent notes and commit messages, explore their psychology deeply by each sentence and each word, then produce comments to specific sentences/words about what you read and share details you noticed. In the end, give a 3-bullet-point summary of their psychological state, their life, their mind, their limitations, their vulnerabilities, the psychological patterns they repeat stupidly.
will do.
Since I use a git repo, this way I’m also adding commit messages/dates to the context window. At some point I used git messages of empty commits to log thoughts
I like to emphasize to the model the importance of figuring out the inner intentions behind single words usage: when journaling there can be lots of information hidden behind word choices, sometimes even more than the journal content itself, and models like GPT 5 are able to uncover some of it.
5. By now the model will have made you notice some things about your behavior based on its internals, let’s now try to validate/connect them with/to research literature and/or spot other patterns that are common in the literature but the model didn’t initially spot on its own.
What else does psychological research/literature suggest about patterns/disorders like these?
6. We have a bunch of insights now, so what to realistically do about them? Let’s see if the model can give hints to answering this question.
I like to keep this real by putting everything on a timeline: everything seems huge and unachievable if you don’t focus on progress and the single first action you can take now.
What are exercises and techniques recommended to {{your name here}} now?? what could be a realistic timeline?
7. 🎉 Congratulations! At this point you have insights and a plan to act on them... after a little common-sense evaluation of them.
8. Another very effective thing I like to do at this point is to ask the LLM to roast me. This makes the model be direct and honest. And, learning to laugh about yourself is a master skill of emotional intelligence.
9. If you have a latex compiler installed on your machine and are using an LLM client that has access to your shell, one cool extra thing to do now is to ask the LLM to write and compile a latex report of the key sentence-by-sentence notes, big picture insights, roasts, and more importantly the plan timeline, so you can leave this with a solid comprehensive PDF to review when needed.
10. Let me know how it goes!
Ending note: probably there is more value in consistent journaling than in the AI analysis itself, so this may be more of an encouragement to journal consistently than an AI guide.
And of course there’s more value in the LLM itself than in the tricks shared in this article.
