How to turn Claude and Obsidian into a personal AI operating system

I pointed Claude at a folder of 17,000 notes and asked it a simple question about my own priorities. It gave me a confident, fluent, completely wrong answer.

Not because the model is weak. Because it had no map. It read whatever fit in the window, filled the rest with a plausible guess, and said it with the same tone it uses when it’s right.

That’s the moment I stopped writing longer prompts and started building three files instead.

Why a bigger context window doesn’t fix this

Claude’s context window is large. It is not your vault.

Drop an unstructured folder into a prompt and say “you know my notes,” and here’s what actually happens: Claude samples what fits, infers the gaps, and answers with identical confidence whether it read the relevant note or never saw it. It isn’t lying. It’s improvising, because nothing told it where to look.

The vault has the signal. Without structure, Claude is working from noise.

The fix isn’t a longer prompt. It’s three files sitting at the root of the vault, read at the start of every single session.

The three files, in order

1. me.md - your portable identity

One markdown file that any AI model can read to know who you are, how you decide, and what you’re working on right now. Not a bio. A briefing document, the kind you’d hand a new contractor on day one.

Builders who’ve been doing this longest keep it deliberately narrow: current projects in one line each, your actual decision defaults (“ship over plan”), communication preferences, hard constraints, and a “current focus” block that gets a weekly rewrite. Everything else - history, old projects, opinions you’re not acting on - goes in a separate reference note, not here.

The point that matters more than the template: this file is model-agnostic on purpose. Point Claude’s own root instruction file at it with a single line - go read me.md first - instead of stuffing your identity into a Claude-specific config. That way when a better model shows up next year, you’re not rebuilding anything. You copy one file into the new tool and you’re at full context in the first message.

2. The vault map - where AI is allowed to look

A vault with 17,000 interconnected notes has no chance of being read whole, and it doesn’t need to be. The map’s job is to tell Claude which ~20 notes are always relevant, which folders are structure (not content), and which folders to skip outright - archives, media dumps, plugin configs.

One structural pattern worth stealing here: organizing the vault into three top-level buckets - one for reference knowledge and ideas, one for anything anchored to a date (daily notes, meetings, journals), and one for active projects and tasks. Three folders, one clear job each. It gives both you and the AI a folder-level answer to “where does this go” before you’ve written a single map entry.

Mistake to avoid: don’t turn the vault map into a living document you touch daily. Build it once per real reorganization. If it needs weekly edits, the folder structure underneath it is the actual problem, not the map.

3. The skill map - behavior, tied to real file paths

A prompt library stores text. A skill map stores what an automation does, which files it touches, and where its output lands - in terms of your actual folder names, not generic placeholders.

This is the file that turns “Claude wrote me something once” into a system that reruns the same way every time, and it’s also where the real leverage compounds - because once you have a skill map, you can group related skills into named systems instead of a flat list.

DAILY BRIEF SYSTEM

Keep the actual skills - the instructions themselves - inside the vault, in this map, not buried in a chat tool’s own memory. It’s fine to use the AI to draft and refine a skill. It’s not fine to let the only copy of it live inside one vendor’s settings page.

What this looks like running

At 6am, before I’ve opened my laptop, one note already exists for the day. It didn’t require me to ask. It pulled the day’s calendar, flagged anything from yesterday still marked open, and pulled the top of the current-focus block from me.md into a three-line “what actually matters today” section.

None of that required Claude to search the vault live. The map told it where the open items live. The skill map told it what format the note takes. me.md told it what “done” means for me specifically.

The part that’s easy to miss: the note has an open space in it - a section you can jot into from your phone during the day, in curly brackets or as an indented bullet. The AI treats that as instructions for the next brief. You’re not just receiving output from the system; you’re leaving it notes back.

The model-swap question, answered honestly

One detail worth being upfront about: pick a default model and reserve the expensive one. The sensible split is a mid-tier model for anything routine - most reads, most rewrites - and the top-tier model only for the tasks where being wrong is actually costly: deep synthesis, research that gets acted on, anything where “close enough” isn’t good enough. The cheapest tier is worth keeping around for pure mechanical work - summarizing an inbound article, tagging - because anything that needs checking afterward wasn’t worth the discount.

The mistake underneath all the others

The whole system fails quietly if the automations live only inside one AI tool’s memory instead of the vault. It works for months. Then a pricing change, a feature deprecation, or just a product pivot, and you’re rebuilding from a memory of how it used to work.

The three files are a bet that plain text you control outlasts any specific AI product. That bet has been correct for forty years of computing. Nothing about this year suggests it stops being correct.

Build it this weekend

One afternoon, three files, in this order:

• Write me.md by hand, honestly - under 600 words, updated weekly on the current-focus block only.

• Build the vault map top-down: what does the AI actually need to navigate to do the work you use it for? Map only that.

• Document one real automation in the skill map before you build a second one.

• Point your AI tool’s own root file at me.md with a single redirect line, so the identity lives outside that tool.

Test it the way you’d test a new hire: paste me.md into a fresh session and ask what your top priority is right now. If the answer is wrong or vague, the file is missing something. Fix the file. Not the prompt.

The AI isn’t the product here. The three files are. The model underneath them is just this year’s layer - useful, powerful, and replaceable without losing anything, if you built it this way from the start.