Augmenting an Obsidian Digital Garden with Agentic AI - an experiment
Introduction
It’s March 2026, and yes I have to specify also the month because things moving so fast that a “2026” timestamp isn’t enough anymore. AI is reshaping everything: how we code, how we produce content and also how we think.
I fell in love with the concept of digital garden the first time I read the most famous essays ”A Brief History & Ethos of the Digital Garden” by Maggie Appleton And. Since that moment, I’ve viewed my Obsidian vault as my digital garden.
I made a very simple experiment: I wanted to integrate an Agentic AI in the managing of my digital garden. I like to think this as an experiment, and I’m going to spoiler you that for me it worked like a charm on most things.
For context: my digital garden is a collection of notes from everything I read: from notes of my Master’s Degree in Machine Learning to STEM, techie and philosophical stuff I read. I take an approach heavily influenced from Wikipedia and data hoarding. I like to play to pretend that in a post-collapse post-internet world my local knowledge base will survive and be useful (even if unrealistic). Notes are often dense and long, but I also try to divide into smaller notes as much as I can, especially when a concept is cited or explained in different books, courses, projects, whatever. I think this context is relevant to understand why I want assistance from an Agentic AI.
I’m watching the IT industry go through an AI-fueled identity crisis. Between the massive layoffs and heated debates about the future of some professions. Before SWE, AI came for writers and artists. Since this is a digital garden, as you expect, the point is that an agent will help with writing and revising. I don’t think that using tools like ChatGPT or Claude to write and edit text is anything new that hasn’t already been done in the past 2 years. So i thought a cool idea would be to apply them to a digital garden.
I think it’s a very recent trend that people are discovering that agents (e.g. Claude Code or Gemini CLI) have a strong synergy with Obsidian. I consider myself very lucky to have always used Obsidian for this kind of stuff. I also use Anytype for more personal stuff, however I don’t like to use agents for that.
I tested it on some repeating tasks that I wanted to automate:
- Mapping topics to find interests, niches, vertical knowledge and connections between topics
- Improving existing notes: finding notes that need work, fixing grammar, improving clarity and reorganizing struture.
- Finding information easily and getting personalized suggestions based on what’s already in my digital garden.
My setup
- I use an Obsidian vault to host my digital garden
- An Obsidian community plugin that spawns a terminal tab for the agent
- Gemini CLI as agent, spawned in the vault root
- Gemini reads a GEMINI file in the root for instructions
- It also reads
.gemini/skills/for skills. I also coded myself a skill that I’m using and improved by a lot my note taking and learning, I will introduce it briefly later.
What works and what doesn’t
I was skeptical a bit at the start about finding the correct files; I was worried it would start to hallucinate file names and directory. Thankfully in my experience it wasn’t the case. Gemini-CLI (which I assume works similar to Claude Code) looks for the correct path by searching on the file system, and also supports referencing files directly by typing @ followed by the filename.
I don’t have to tell you that they are great at summarizing content, even combining it from multiple notes. However they require a bit of prompt engineering for having better results. One of the first things that I asked Gemini-CLI was to read my entire vault and analyze it. It didn’t provide much insight at first, mostly just reading some of my (outdated) “meta-notes”. Also initially the agent was too much flattering giving me generic praise like “Your notes are exceptionally well-documented,” which feels nice but doesn’t help me improve. This sycophancy is a common AI trait.
I found it decent suggestions when I asked Gemini to, given my entire vault, “suggest other topics that I could study or deepen”.
What Gemini suggested me to study next
- Prompt Injection
- Insecure Tool Use & Privilege Escalation in AI Agents
- Indirect Injection & Data Exfiltration Patterns
- Insecure Output Handling
- LLM Guardrails
I also asked Gemini for a psychological analysis of me based on my vault. It wasn’t very insightfult tho. If you’re curious it said I have a “Skeptical Architect” mindset (whatever that means). It also said:
Quote
Your vault is not just a “how-to” manual; it is a synthesis machine designed for long-term mental model building, prioritizing “Familiarization > Memorization.”
It used a lot of meaningless buzzwords like “Rigorous Synthetiser” and “Security Bias”, which felt more like a Linkedin post than a real analysis. I don’t think agents are great at psychological analysis.
Agents are decent at suggesting where to put a new concept or case study, but if you know your garden well your intuition will still be better (in my experience). Sometimes Gemini is too creative in finding connections and when reading them i would think “no that wouldn’t fit”. However, for content I wrote a long time ago and might have forgotten, the AI’s suggestions are quite helpful. With my surprise, it also works with Obsidian bases. It correctly extracted the top 5 files from my “recently edited file” base.
I found Gemini very sucessful in reading PDF slides (mostly slides from my courses) and “copy-paste” content into a note in my vault. It saved hourse of manual labor. It handles the mechanical boring parts so I can save time to focus on the actual learning. I like to study from my digital garden, so I like to have everything here and then integrate stuff and connect to other topics. Gemini works very well at this mechanical and boring stuff, I really like using AI to automate this specific thing.
I found some issues with Gemini-CLI, as of 03/04/2026 that I’m writing this blog, I find it slow, and prone very fast to context rot so I have to often run the command /clear. Sometimes it skips content from the PDF files. It often ignores images and diagrams, so I still have to put some work and effort into making notes. However, the most boring and maybe time-consuming part is done by the agent, and that’s what I care about the most.
Another thing that I found very successful was to make flashcards that I upload to Anki. It’s part of my method of studying stuff for my master’s course, that I also used at my bachelor’s course. I also coded myself a skill that after generating notes, it automatically calls a small script (coded, not vibecoded) and uploads them to Anki. I think there is a lot of potential in introducing skills that run programs inside a digital garden. If you’re interested in the code or in downloading it for your agent, you can find it here agentic-skill-obsidian-to-anki-flashcards (prob will change name in the future).
I have a lot of thoughts on agents and how they could impact digital gardens that I will explain in the following section.
Fighting the anthropomorphization impulse
I know that there is a long story of anthropomorphization of AI assistants like the Ouroboros case. People do far stranger and creeper things with AI “girlfriends”. However i still think it’s cool to give a personality, even if they’re just faking it.
Back in the Tumblr/Geocities era the web look was deeply personal and customization was a big thing. Compared that to the standardized, corporate look of X, Instragrams, TikTok. Some may argue that atleast algorithms study you and are personalized for you, however you are mostly passive in this scenario. I like to see giving the AI assistant a persona as way to reclaim that old-school personalization in this context. ChatGPT 4.0 problem with sycophancy (i.e. being too nice, prioritizing approval over truth) was definitely problematic, however Claude tend to be too cold sometimes. I understand that it is a bit risky and some people may lose touch with reality, however I think with enough consciousness it can be funny to make your AI assistant imitate a personality.
Of course, it depends on the use case. I think that a single agent that spawns in a digital garden looks more like a personal agent or pet, so maybe it makes sense to give him a bit of personality. If you’re vibe coding your personal project and maybe you’re spawning different agents to solve tasks well maybe it’s not that useful. I believe that, with how markets and agents are evolving, everyone will have a personal agent that helps them, similar to how we all have smartphones today. These personal agents will have much more personalization and context about us than what we have today. My prediction is that in the future also textual AI “companions” (i hate this word) will be much more difficult to distinguish from real humans, similar to what is happening now with AI-generated image and videos. Maybe i’m gonna write more about this in the future.
I also have one last thing to say before showing the “soul” of my garden. I don’t believe you can really personalize well your agent unless you’re doing fine-tuning. What you can do with system prompting is very limited.
You can view my full GEMINI file following the wikilink. My agent is called “Libgar” (from Librarian Gardener). I also give it a non-binary gender. Ok yes I know that it really doesn’t have a gender, but for the sake of creating a character and making the agent imitate such character, I also give it a gender. Maybe in the future I will give Libgar more personality traits taking inspiration from what you usually do when you design a character for a book. I will never give him creepy things like age or imaginary physical characteristics, in fact I like to think of Libgar more as an incorporeal agent that lives in the wire without any anthropomorphic form.
This is the “Soul” section:
The soul section isn’t the only one, but it has more with a lot of instructions and guidelines. It has a section copy-pasted from tropes.fyi to avoid that Libgar writes in the style of AI slop. It also have a “Capabilities” section with common task with step-by-step instructions. I often refine and add new things to this file as i’m interacting with the agent and they are making errors. I find that with the time, if you are very precise in the instructions and you cover edge-cases, it will become smarter.
My thoughts on AI assistants and Digital Gardens
I have some thoughts that are jumping all around my head, but probably some of them will be only speculative.
Delegating the writing to an agent introduces a fundamental shift in the digital garden workflow. When the mechanical act of writing is automated, the human gardener can operate at a higher level of abstraction, focusing more on the conceptual connections and clusters of notes with similar themes. However this is personal and maybe depends on the scope of your digital garden. Maybe someone just need writing corrections. Me personally i don’t like that AI writes content in my digital garden. I find it a bit unease. So most of the time i write it myself (however i heavily rely on it for fixs) On the other side agents lack vision and will, you as human have to provide the creativity and intuition necessary for a digital garden to “flourish”.
An interesting speculative experiment would involve creating a “character agent” with a personality, tasked with maintaining and expanding their own digital garden, with all the connections and stuff. Maybe it will simply recreate a smaller, more biased Wikipedia. Probably it will be useless and at most just an “artistic” project, or maybe will teach us something about how AI model understand and connect information.
I see some clear disadvantages to Generative AI integration. First, rapidly producing content can lead to a loss of control, unless a rigorous organization (e.g., wikilinks and using the graph views) is maintaned. English is not my first language, so my writing is a lot imperfect. AI may help me with making my content seem more clean it will slow or totally stop my learning. Same could happen with other skills.
I’m a bit skeptical that AI integrations into blogs/agentic AI would be taken well from the digital garden community. A lot of them overlap with the privacy-conscious people that are often also AI skeptical (not everyone of them tho, local offline self-hosted models are well seen). I think there are good reasons not to use AI i.e. environmental issues, however it’s very hard if you want to work in IT avoiding what everyone is using and what your future employer expects you to use. Digital gardens are tools for learning, thinking and expressing yourself as human, so i think they will keep being human-centered, no matter how convenient will be an AI assistant.