Why a note-taking app was no longer enough

I use Obsidian to collect notes from university, projects, books and podcasts. For a long time the vault was mostly an archive for myself: I wrote something down and found it again later, as long as I knew where to look.

Working with AI agents shifted the problem. Every new session with Claude Code or Codex started without any knowledge about me. I explained again what I study, which projects I am working on, how I write and where new notes belong. That information already existed. It just was not where an agent would find it.

So I rebuilt the vault. It should stay readable for me and at the same time be the foundation my agents start from. By "operating system" I mean exactly this layer: a folder structure that expresses relevance, context files about me, rules for handling files, and skills that describe recurring workflows. An agent that starts in this folder finds that framework already in place.

The vault now holds about 930 Markdown notes. At that size, what matters most is that an agent finds the right note and knows what it is allowed to do with it.

PARA as a shared structure

The basic structure follows Tiago Forte's PARA method. PARA stands for Projects, Areas, Resources and Archives. Projects have a concrete goal and an end. Areas are ongoing responsibilities without an end date, resources are reference material for later, and the archive holds whatever is not active right now.

I added three folders to the four categories and numbered them: 01 Inbox, 02 Projekte, 03 Bereiche, 04 Ressourcen, 05 Daily Notes, 06 Archiv and 07 Anhänge. The folder names are German because the vault is. The inbox is the entry point for anything that does not have a place yet. Daily notes record what was decided on a given day and what is still open. Obsidian automatically stores pasted images and PDFs in the attachments folder.

For agents, this structure has a practical advantage: the folder already says something about a note's status. A file in 02 Projekte belongs to ongoing work. A file in 06 Archiv is finished and should not be treated as the current state. When I ask what I am working on, an agent can read the active projects and the latest daily notes instead of searching the whole vault.

On top of that come a few simple conventions: frontmatter with tags, status and date, file names in normal spelling, and wikilinks between related notes. Areas and resources are always folders because they grow over time. A project starts as a single file and only gets a subfolder once it needs several files. These rules live in the vault itself, so that the agents and I file things the same way.

Context every session finds

The top level of the vault contains an AGENTS.md. Codex or Claude reads the AGENTS.md as soon as I start a session in the vault folder. The file is short: a paragraph about me, a description of each folder, the rules for filing and naming, and a few routines.

The routines turn the description into a workflow. At the start of a session the agent checks the inbox and offers to sort new entries. For questions like "Where did I leave off?" it reads the last two or three daily notes and the active projects. At the end it offers to write an entry for the day and to save new insights as notes. That creates continuity between sessions without me having to prompt it every time.

The hidden .Kontext folder goes into more detail. It holds files about me, my services, my target audiences, my branding and my writing style. Obsidian hides folders whose names start with a dot, so they stay out of the way in daily use. Agents can still read them, and AGENTS.md states explicitly when they are needed: when writing texts, emails or proposals.

The writing-style profile shows how precise these files have to be. It is based on sent emails, my own correspondence in the vault and my input in conversations. It separates texts I wrote myself from notes that were clearly produced by an AI. Without that separation, an agent would use its own earlier output as evidence of my style. I also use the profile for texts like this one.

I adopted this setup from a YouTube video by Julian Ivanov (in German).

Skills: how agents work with the vault

Context files describe who I am and how the vault is organized. Skills describe workflows. A skill is a folder with a SKILL.md: a short description of when it is needed and instructions that the agent only loads when the task calls for them. That keeps the base context of every session small.

The vault contains six skills, stored once in a folder for Claude Code and once in a folder for Codex and other agents. Four of them come from Steph Ango's public obsidian-skills repository: Obsidian Markdown with wikilinks, callouts and properties, Bases for table views, JSON Canvas and the Obsidian CLI. They teach agents the formats Obsidian understands beyond plain Markdown.

I added two skills for my own workflows. The first sorts unfiled notes into the right PARA folders based on their content. The second turns a book into a structured note with a summary, core ideas and quotes, plus a character map as a canvas for novels. A Bases view then shows the book notes as a table with author, status and rating.

In daily use the effect is small but noticeable. I no longer have to explain what a callout looks like or that related notes should be linked. An agent that creates a new note writes it in a format Obsidian renders correctly and follows the rules from AGENTS.md.

An assistant that maintains the vault

Local agents only work while my laptop is running. For recurring tasks I therefore use Hermes, my personal assistant on a virtual server, which I also describe in the Code factory case study. The vault lives in Google Drive. Hermes accesses it through the Drive API and works with the same folders and rules as the local agents.

Several workflows run on a schedule. Every hour between 8 a.m. and 10 p.m., inbox upkeep clears the inbox folder: notes get a descriptive name and move to the matching PARA folder. If a note contains a task, it can become a reminder or a ticket. In the morning Hermes compiles priorities from my calendar and open tasks and checks the folder structure. On Sundays it writes a weekly review with completed items, open topics and a plan for the next week.

Knowledge also enters the vault through fixed routes. Readwise syncs highlights from books and articles, and podcast clips come in through Snipd. When I send Hermes a YouTube video, it saves a summary as a note. Handwritten notes from GoodNotes are converted to Markdown. After that, everything is searchable and readable for agents.

Hermes now has about 110 skills. Each one is mirrored as a Markdown file into 03 Bereiche/Agent Skills, grouped by category and listed in an index. That way I can look up in the vault what my assistant can do and which rules it follows. The vault also documents the agents that work with it.

My degree in the vault

The benefit is clearest at university. Under 03 Bereiche/Studium, every course this semester has its own folder with one subfolder per week. Lecture slides, exercise sheets and my own notes land in the matching week folder automatically. A separate note maps the events in my university calendar to the course folders, including cases where the calendar title and the folder name differ.

The lecture follow-up skill builds on that and runs on the server every evening at 6:15 p.m. It checks which classes took place that day, reads the week's material and creates three files per course: a summary whose core can be taken in within about a minute, study exercises with answers that can be revealed one at a time, and flashcards as Markdown for RemNote, at most 20 per course and week.

The exercises follow the structure of my study plans from last semester: short check questions, definitions, explanation questions in the style of the Feynman technique, and application tasks. If I fill in answers in an earlier week or mark something as uncertain, the next run picks up that gap in a short review. It explicitly does not count empty fields as a knowledge gap.

I added one rule after the first runs: without a new or changed file in the week folder, no new output is created. Before that, the run had produced files on a day with a class but no new material. Now a small script compares the source files with the last saved state using SHA-256. Earlier outputs stay unchanged, and every new version gets its own timestamp. I still import the cards into RemNote myself.

Versioning and clear write rules

When several agents work in the same files, I need a way back. The vault has therefore been a Git repository since February 2026. The Obsidian Git plugin stores backups as commits in a private GitHub repository, 85 so far. That makes it possible to trace when a note changed and to undo a mistake.

Git does not replace rules. AGENTS.md therefore says that agents ask before deleting or overwriting anything, briefly explain why they move files, and only move projects to the archive when I tell them to. The lecture follow-up skill creates new versions instead of replacing old files. The same rule applies to Hermes in the vault: it deletes nothing.

The Git configuration separates content from local settings. Session data and app settings stay out of the repository, while the skills in the vault are deliberately versioned. That also makes it traceable when the instructions to the agents changed.

What the system does not solve

A vault that agents work with is only as good as its content. Notes go out of date. A project note might describe a plan that was implemented differently long ago. Current sources such as code, the calendar or the files themselves therefore take precedence, and older notes are a hint that needs checking.

Part of the content is now generated by AI itself: summaries, weekly reviews and study material. Those files are marked, for example with a frontmatter field naming the skill that produced them. They can still contain mistakes. A generated summary also does not take away the work of actually learning the material.

Privacy is the second limit. When an agent reads notes, that content goes to the respective model provider. The hidden context folder hides files in Obsidian; it does not protect them from that access. Every integration therefore needs a decision about which areas it really requires. Credentials do not belong in the vault.

I have no measurement of how much time the system saves me. I have to explain far less, and sessions get to the actual task faster. That is an observation from my daily work, not a reliable number. Maintenance is part of the picture as well: rules, skills and folder structure have to fit the way I work and change along with it.

What I take away from it

For me, the most important lesson is that context needs its own infrastructure. A strong model does not help much if it does not know what I am working on. With Markdown files, a fixed order and clear rules, that context is readable for me and for any tool that can work with files. If I switch agents, the vault stays.

That matches what I learned from the code factory: handoffs need a fixed place. There it is tasks and test results between agents. Here it is knowledge about me and my work between individual sessions.

Next, I want to understand better which notes agents actually use and where they run into outdated information.

For companies and clients, the same question sits underneath my own setup: what context does an agent need, where does it live, and what is the agent allowed to do with it? My vault is one concrete answer for one person. In a team, permissions, shared data and responsibilities would come in as topics of their own.

Sources