Create a Claude account and pick a plan
5 minGo to claude.ai and sign up with your email. Then choose a plan — Pro is enough to get going. You can change plans any time.
A setup guide · Obsidian + Claude
Step by step, from a blank laptop to a private library of every Word and Google Doc you’ve ever written — one you can search, sort by date or topic, and ask questions of in plain English. No coding, nothing to install beyond two apps. About an hour of setup, then Claude does the heavy lifting. Every prompt on this page has been run end-to-end on a test archive.
Read first · 3 min
You’re setting up three things that work together. Once they’re connected, you mostly just talk to Claude.
A free app that shows a folder of plain-text files as a beautiful, searchable library. This is where you'll read and search your writing.
The AI. It runs on your laptop, reads that same folder, and does the work — converting files, organizing, tagging, answering questions.
The folder itself, pre-organized with numbered subfolders and instructions that teach Claude how to file and connect things. It's yours; it lives on your computer.
Phase 1 · about 20 min
Nothing here needs a terminal. Download, open, sign in.
Go to claude.ai and sign up with your email. Then choose a plan — Pro is enough to get going. You can change plans any time.
This is the app you’ll talk to. Download it, drag it to Applications, open it, and sign in with the account you just made.
Once you’re in, look at the top of the window: there are tabs like Chat, Cowork and Code. You’ll be using Code — that’s the version of Claude that can open a folder and work inside it.
Free, no account needed. Download, drag to Applications, open it once. If it offers Obsidian Sync or a paid plan, skip — you don’t need them.
The vault uses two small helpers that come with your computer: Python (for the conversion) and git (the undo history). On a Mac you don’t install anything. But the first time Claude uses one of them, macOS may show a pop-up:
On Windows, install Python from python.org (tick Add python.exe to PATH on the first installer screen) and Git for Windows — the Claude app’s Code tab needs it. If either goes wrong, don’t fight it: in step 8 you can tell Claude “install it for me.”
Phase 2 · about 20 min
You’ll download a ready-made folder, put it in Documents, then open it in Obsidian and in Claude. One thing to know up front: the download is a folder with a folder inside it. Obsidian opens the inner one; Claude opens the outer one. Both steps below say exactly which.
Open it and look inside. You’ll see:
Leave everything exactly where it is.
You should now see the numbered folders down the left side. Click 00_HOME → Start Here to have a look. Don’t worry about understanding it all yet.
First, a quick safety habit that the kit’s own author recommends whenever you download something onto your computer:
I'm not technical and this is my first time here. Please read through this whole folder and tell me, in plain English, what it is and what it does. Is there anything in it that could harm my computer or send my files anywhere? Keep it short.
Read the answer. Then start the built-in setup by typing exactly:
let's go
Claude now walks you through a short conversation, one question at a time. Here is what it will ask and what to say. Use your own words — these are just shapes:
I'm going to bring my whole archive in as one big batch in a moment, following a guide. Please skip the single-file step, set up git now, and tell me when you're ready for the batch.
Claude sets up git (the invisible undo history for the whole folder — it means nothing you do later is unrecoverable) and says it’s ready. That’s the end of setup.
Phase 3 · 15 min of clicking, then waiting
Goal: one folder on your Desktop called My Writing with everything in it. It does not need to be tidy — Claude sorts it.
You are collecting copies, not moving anything. Your originals stay exactly where they are. It doesn’t matter how many there are — 40 or 4,000 — and it doesn’t matter how messy the names or folders are.
Google Takeout packages up your whole Google Drive as a zip. It’s the reliable way to get years of Docs out at once.
The vault has an “in tray” folder called To Process. Everything Claude works on starts there. It’s buried a few folders deep, so let Obsidian open it for you instead of hunting for it:
Phase 4 · Claude works, you supervise
Four prompts, in order. Copy each one into Claude (Code tab, Writing Vault selected), press Enter, and let it run. Read what it reports back before moving to the next. Every one of these has been run end-to-end on a test archive; they work without installing anything.
One setting first: next to the send button there’s a mode selector (it probably says Manual). Switch it to Accept edits for this phase. Claude will still ask before running commands, but won’t stop for every single file it writes — otherwise you’d be clicking Allow hundreds of times.
Before converting anything, have Claude count what’s in the inbox and compare it to the number you wrote down in step 9. This catches a half-finished copy or a folder dropped in the wrong place before an hour of work is built on it.
Look inside vault/01_INBOX/To Process/My Writing and tell me what's there. I want: - for the Word folder and the Google folder separately: how many .docx files, how many .doc files, and how many other files (by type) - the total number of Word documents overall - the oldest and newest file dates you can see Don't change, move, or convert anything yet — just report.
Does the Word folder total roughly match your Finder count from step 9? A few off is fine (Finder counted the folder itself and any stray files). If it’s way off, the copy probably didn’t finish or the folder went somewhere else — tell Claude: “I expected about 2,300 Word files. Help me figure out where the rest are.” If it matches, carry on.
Word and Google files are locked-up formats. This asks Claude to turn each one into a plain-text Markdown file (which Obsidian can search) while keeping every original untouched — using only what’s already on your Mac. It’s the longest-running step; the batch reports let you watch it move.
In vault/01_INBOX/To Process/My Writing there are Word documents (.docx, and possibly older .doc files) and a Google Takeout export of my Google Docs (also .docx). This is years of my writing. Please: 1. Convert every document to a Markdown (.md) file. Use only tools already on this Mac — textutil (built in) converts .doc and .docx, and you can write a small Python script if you need to. Do NOT install Homebrew or anything that would ask me for a password. If you truly can't proceed without installing something, stop and tell me what and why first. 2. Do a dry run first: show me the plan (how many files, where they'll go) and wait for me to say "go" before moving anything. 3. Keep each file's original filename as the title (just change the ending to .md). Keep headings, paragraphs AND line breaks — some of these are poems, and a line break inside a stanza must stay a line break (in Markdown that means two spaces at the end of the line). Test this on a poem before doing the whole batch. 4. At the top of each .md file add frontmatter with: title, date, date_source, word_count, source (word or google), and original_path. For the date: use a "Written ..." line or a clear date near the top of the text if there is one; otherwise the document's own date. In date_source just say in plain words which you used (e.g. "Written line in the text" or "file date"). 5. File the .md files in vault/02_SOURCES/Writing/, in subfolders by year. If the only date you can find is a file date from this year (that's just when I copied it), treat it as undated and put it in vault/02_SOURCES/Writing/undated/. If two files would get the same name in the same year, add (2), (3) — never overwrite. 6. Move the untouched originals into vault/02_SOURCES/Writing/_originals/ (keep their folder structure) so nothing is lost. Never modify or delete an original. 7. Anything that isn't my writing (spreadsheets, PDFs, images, web pages) — move it to vault/01_INBOX/To Process/_skipped/ and list it for me. 8. Work in batches of about 25 files and give me a one-line progress report after each batch, like: "Batch 3 done — 75 of 2,314 converted." (If I only have a few dozen files there'll only be a couple of lines — that's fine.) 9. Save the script you write in tools/ so we can run it again later. If I run it again with new files in the inbox, only process what's still in the inbox — never redo files already in 02_SOURCES/Writing/. When you're done, tell me: how many files you converted, how many you skipped (and list them), any you couldn't read, and confirm the My Writing folder is now empty. Then save an undo point with git.
Now you have hundreds of clean files. This creates one page that lists all of them with dates, topics, and a one-line summary — the table of contents for your life’s writing.
Read every file in vault/02_SOURCES/Writing/ and build a page at vault/00_HOME/Writing Index.md. Make it a table with one row per piece: title (linked to the file), date, word count, topic tags, and a one-sentence summary of what it's about. Sort newest first. Also add a "tags:" list to each source file's frontmatter — that's fine, it's metadata at the top of the file, not my text. Use a small, consistent vocabulary of about 15 tags total: each piece gets its main subject plus whatever else genuinely applies (usually 2–5). Reuse the same tag names rather than inventing new ones. When you're done, show me the full list of tags with how many pieces have each one, and save an undo point with git.
Open 00_HOME → Writing Index in Obsidian. This is your library catalogue. It’s a normal file — you can edit any line.
This is where it stops being a filing cabinet. The kit ships with skills for exactly this: diarize builds a page about one subject from every source that mentions it; emerge looks for patterns nobody named.
Using the Writing Index and the tags, tell me the 10 subjects I write about most often across all my writing, with a rough count and date range for each. (Subjects — not recurring phrases or people's names, unless a person really is the subject.) Then, for the top 5, run the diarize routine (vault/.claude/commands/diarize.md) on each one to build a canon page in vault/04_CANON/Practice/Themes/ — the vault's usual place for themes. Skip the web-search step; these are my private essays. Each page should describe the theme, how my thinking about it changed over time, and quote or link the specific pieces where it shows up. Every claim should link back to a source file. Finally, run the emerge routine (vault/.claude/commands/emerge.md) once and tell me what patterns you found that I might not have noticed. Then save an undo point with git.
Phase 5 · everyday use
From here on there are two ways to look through your writing. Use whichever suits the moment.
Open Claude (Code tab, Writing Vault) and ask in ordinary language. It reads the relevant files before answering, and it can make new pages for you. Some starters:
Find everything I've written about my father. Give me the passages, with a link to each file and its date. Which pieces from 2018–2021 are about work? Sort them by date and give me one line on each. Show me my ten longest pieces and my ten shortest. I think I've written about "leaving" a lot. Am I right? Where, and how did the way I wrote about it change? Make a page that groups all my writing into 8 themes, with the pieces listed under each. Pull the best three paragraphs I've ever written about New York into one file so I can reread them. I want to write a new essay about X. What have I already said about it, and what haven't I said yet?
Two habits that keep the vault healthy, straight from the kit’s authors:
At the end of a working session:
Write a recap of this session to the log — the recap routine (vault/.claude/commands/recap.md). Then save an undo point with git.
Claude writes a dated note of what changed at the top of 00_HOME → Log, so next time you can ask “where was I?”
About once a week:
Run the vault health check — the vault-lint routine (vault/.claude/commands/vault-lint.md). Tell me which findings are real problems versus template noise, and fix the 3 real ones. Then save an undo point with git.
It checks for broken links, duplicates and stray files. The first run always finds a lot; most of it is scaffolding, which is why the prompt asks Claude to sort real from noise.
Phase 6 · read when needed
It can see the whole vault, so it’s a better troubleshooter than any help page. Say exactly where you are, even if that’s “I don’t know what to ask.”
I'm stuck and I'm not technical. Explain to me like I'm brand new: what state is my vault in right now, and what should I do next? Give me 2 or 3 options and tell me which one you'd pick.
Something went wrong and I want to undo what just happened. Use git to show me what changed in the last hour, and put it back the way it was if I say so.
The vault is a folder on your laptop; nothing syncs anywhere unless you set that up. When you ask Claude something, it sends the relevant files to Anthropic’s servers to think about them and then they’re gone — same as pasting text into any AI chat. The kit’s own rule of thumb:
Good first-week prompt: “Scan my vault for anything that looks like a password, account number or private ID and tell me where it is.”
In the outer Writing Vault folder are three guides written by the kit’s authors — double-click to read, or ask Claude “show me RECIPES.md”: RECIPES.md (24 things to try, each with input → prompt → result), Standard Operating Procedure.md (daily and weekly routines), and Data Sources to Gather.md (what else you could bring in — photos, email, old chats).
Phase 7 · what it can and can’t do
Once the writing is in, this is what to ask for, what to expect back, and where the edges are. Everything here was tried on a test archive of essays and poems.
Obsidian doesn’t read or sort anything — it’s the window and the search box. Claude does the reading: it can open every piece, decide what it is and what it’s about, make folders, add tags, write pages, and answer questions. Anything Claude organizes shows up in Obsidian a second later, and you can drag things around yourself if you disagree with a call.
The one design idea to hold onto: folders for kind, tags for theme. A piece is one thing (a poem) but about many things (your mother, Brooklyn, leaving). Folders can only hold a piece in one place; tags can mark it with ten. So kinds become folders, and themes become tags plus theme pages.
Best run right after Prompt 1 (before the index), but it works any time — Claude updates the links if pages already point at the files. Change the list of kinds to whatever fits your writing.
Sort my writing by kind. Read every file in vault/02_SOURCES/Writing/ (not _originals) and decide what each one is: Poem, Essay, Journal, Draft, or Other. Rules: judge by the writing itself, not just the filename. If a piece is a poem that's also a draft, it goes in Poems (kind: Poem, plus a "status: draft" line) — kind wins over status. Anything you're not sure about, tell me and take your best guess. Then move each .md into a subfolder by kind — vault/02_SOURCES/Writing/Poems/, Essays/, Journal/, Drafts/, Other/ — keeping the year folder inside (e.g. Poems/2015/Kettle.md). Add a "kind:" line to each file's frontmatter. Leave _originals exactly where it is. If any page in the vault links to a file you moved, update the link. Also update the import script in tools/ so that next time I bring in new files it sorts them by kind the same way. When you're done, show me the counts per kind and the list of anything you weren't sure about, then save an undo point with git.
In testing, Claude sorted 44 pieces into Essays 34 / Poems 5 / Drafts 5 in one pass, explained its reasoning (poems: short, lineated, stanza breaks; essays: prose paragraphs), and listed four borderline calls for the writer to decide. That “not sure” list is the part to read.
Prompt 2 already tags every piece and Prompt 3 builds theme pages for the top five. From there:
All of these work today. Claude searches for likely words, reads the candidates, and answers with links you can click straight into the file.
Find everything I've written about my mother. She's "Mom", "Ma", "my mother", and sometimes just "the kitchen". Give me the passages, each with a link to the file and its date, oldest first. Which pieces from 2012 to 2016 mention the apartment on Graham Avenue? I wrote a paragraph once about a letter I never sent. Find it — it might be in an essay or a poem. Show me every poem that mentions light. Which image comes up most often across the poems?
List my essays longest to shortest with word counts. Which years was I most productive? Give me a table: year, number of pieces, total words. I have three versions of "The Pitch". Show me the differences between them and tell me what changed in how I told it. Group all my journal entries by season and tell me what I tend to write about in winter versus summer.
How did the way I write about leaving change between my twenties and now? Quote the evidence. Which people appear across the most pieces? Make a page for each of the top five with every place they show up. Read all my drafts. Which three are closest to finished, and what would each need? I want to write a new essay about running. What have I already said, and what haven't I said yet?
Pull my best three paragraphs about New York into one file so I can reread them together. Make me a chronological reading list of everything about my father — a page with links, one line on each. Draft a 300-word bio of me as a writer, using only what's actually in the vault. Cite the pieces you drew from.