<aside>
AI news for humans. Once a week.
Plain English. First-person tested. Built for people who feel behind on AI. Each issue ships with a free pack like this one: the actual files, prompts, and shortcuts to try the tool yourself.
New here? Welcome. → Subscribe to KUWAI
</aside>
A neural network of every thought, conversation, and idea from the last five years of my life. This is how it's built.
This is the companion to Day 01 of Finding AI tools that are actually worth it, the weekly Reel series. The Reel's the review. This page is the how-to.
I tested Obsidian as the vault layer for a personal AI knowledge base, paired with Claude Code as the agent that maintains it. Plain markdown files on disk, a wiki Claude builds and re-reads as it learns, all your business context flowing in automatically every morning.
If the review made you want to try it yourself, the setup starts here.
One paste, plus an optional install that makes it smarter.
Step 0. Install the BYS skill (optional, recommended).
📦 Download the skill (zip):
A Claude Code skill that researches every documented way people have done a thing, filters through your specific situation, and gives you a recommended path with the alternatives flagged. Built for "I'm about to start X" moments where a naive Google search or a one-shot ChatGPT reply isn't enough.
Four stages: The Sniff (it asks about your situation), The Prowl (disciplined multi-source research), The Pounce (synthesis), The Way (the personalised plan).
To install: unzip the download above, drop the before-you-start/ folder into ~/.claude/skills/, then start a Claude Code session. ~30 seconds.
Having BYS installed makes the master prompt below smarter. Claude pulls live research where Maran's documented path needs adapting to your stack. You'll also have BYS for any future project where you're starting something new. It's the same skill, just pointed at a different challenge.
Skip Step 0 and the prompt still works on its own.
Step 1. Paste this prompt.
It interviews you first (your tools, your stack, your vault name, your style), then walks you through the entire setup end-to-end. Pause-points built in.
Help me set up Obsidian + Claude Code as a personal AI knowledge base, end to end. Walk me through it in one session. Where you can do the work yourself, do it. Where I need to click in another app or sign in to something, tell me clearly and pause for confirmation.
== CONTEXT ==
I want a personal AI knowledge base built on Obsidian (markdown files on disk), maintained by Claude Code. End state: my whole stack flows in automatically, Claude has my full operating context across every session, and I have a wiki layer that gets richer as more sources are ingested. The architecture is documented at solstudiosdigital.notion.site/01-obsidian.
== STAGE 1: INTERVIEW ME FIRST ==
Before doing anything, ask me about my situation. One question at a time, wait for my answer before the next:
1. What do you want to call your vault? Suggested defaults: "Brain", "Vault", "Knowledge", or something personal. Maran's example was "Big Brain". Pick whatever fits you.
2. What OS are you on? Mac, Windows, or Linux. This determines launchd vs Task Scheduler vs systemd for sync scheduling.
3. Confirm you have Claude Pro or Max. Required for Claude Code; Pro is fine to start.
4. Cross-device sync or single-machine? Obsidian Sync is $4 AUD/mo. iCloud, Syncthing, or a self-hosted alternative all work too.
5. List the tools you actually use that you want me to sync into the vault. Tell me YOUR stack, not someone else's. For each, I'll later build you a sync script if there's an API. Examples to spark thinking: Notion, Slack, Gmail, Google Calendar, Google Drive, Todoist, Apple Notes, Granola, Readwise, Linear, Things, Discord.
6. Tell me about you. Your name, what you do, any voice/style preferences for your CLAUDE.md (casual or formal, emojis OK or banned, language quirks I should respect, things you hate seeing in writing).
7. Anything you DON'T want me touching? Sacred zones. Default is raw/ and .obsidian/. Add more if relevant.
Wait for all 7 answers. Reflect what you heard back to me in 3-4 lines. Confirm I'm happy with the picture before moving on to Stage 2.
== STAGE 2: THE PLAN ==
Walk me through the 10 steps below in order. Pause after each for me to confirm or redirect. Where the step is purely terminal/code, just do it. Where it's UI-based, give me click-by-click instructions and wait.
1. INSTALL OBSIDIAN + CREATE VAULT (UI, 15 min). Tell me to download from obsidian.md, install, create new vault using the name I gave you in Stage 1.
2. SYNC + E2E ENCRYPTION (UI, 10 min, optional cost). If I said yes to cross-device sync, walk me through Settings → Core plugins → Sync, subscribe at obsidian.md/pricing (Standard tier), enable end-to-end encryption with a password I save in my password manager. Skip this step entirely if I said single-machine.
3. FOLDER STRUCTURE (you do it, 5 min). Tell me to toggle in Obsidian Settings → Files and links: "Automatically update internal links" ON, default new notes = inbox, default attachments = attachments, wikilinks ON. Then run from inside the vault: mkdir -p raw/clippings raw/imports wiki/concepts wiki/entities wiki/sources reports projects daily inbox templates attachments
4. AI SURFACES (mixed, 15 min). Confirm Claude Code in terminal (run `which claude`). Tell me to install Claude Desktop if I don't have it. Walk me through Claudian (sidebar chat in Obsidian): install BRAT, add the beta plugin from github.com/YishenTu/claudian, paste my `claude` binary path into Claudian settings. WARN ME: don't install plugins that ask for an Anthropic API key. Anthropic banned third-party OAuth in April 2026.
5. WRITE CLAUDE.MD (you, 30 min). Use the answers I gave you in Stage 1 to draft a CLAUDE.md at the vault root. Under 150 lines. Cover: who I am, what the vault is and isn't, folder map, sacred zones, autonomy rule (default: "act first, log after; only confirm on more than 10 file moves, renames, deletions"), age-weighted ingestion (under 30 days = full depth, 30 days to 1 year = standard, 1+ years = seed-only with corroboration flag), my voice/style rules. Don't write "ALWAYS" or "NEVER" in caps. Show me the draft, let me edit, save when I confirm.
6. WIKI ENGINE (you, 10 min). Run: claude plugin marketplace add AgriciDaniel/claude-obsidian. Then: claude plugin install claude-obsidian@claude-obsidian-marketplace. That gives /wiki, /wiki-ingest, /wiki-query, /wiki-lint, /save, /autoresearch, /canvas. Run /wiki to scaffold the meta layer.
7. SMART CONNECTIONS + LOCAL REST API + MCP TOOLS (UI, 10 min). Walk me through Obsidian Community plugins: install Smart Connections, Local REST API by Coddingtonbear (turn on "Enable Non-encrypted (HTTP) Server", note the API key), MCP Tools by jacksteamdev. Then update Claude Desktop's MCP config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac, equivalent on Windows/Linux) with the obsidian MCP server using that API key. Restart Claude Desktop. Test with "list files in my vault". Use jacksteamdev/obsidian-mcp-tools, NOT MarkusPfundstein's mcp-obsidian (it lacks Smart Connections hooks).
8. DATA SYNCS (mixed, 1-2 evenings). Use the tool list I gave you in Stage 1. Don't try to do all of them in one go.
8.A First, ask me which tool I want to set up first. Suggest the lightest one to build confidence.
8.B For each tool I want to sync:
- If it's UI-only (Apple Notes, Web Clipper, browser extensions): walk me through the install + import.
- If it's API-based (Notion, Slack, Google services, Todoist, Linear, etc.): I'll create the integration or OAuth credential, give you the token. You write a Python sync script + scheduling job (launchd on Mac, Task Scheduler on Windows, systemd on Linux). First sync may take time; after that, daily.
- If it's local-only (Mac dictation, browser bookmarks): point a script at the local data store.
8.C Stagger sync schedules so they don't fight each other (15-min increments).
8.D ALWAYS warn me to do a privacy pass after any one-shot import. Smart Connections embeds everything; sensitive content (passwords, tax records, medical) should be deleted at source before vectorising.
9. CALIBRATE WIKI (you, 30 min). Pick 2 real notes I care about. /wiki-ingest each. Show me the result. Ask if it over-extracted or under-linked. Adjust CLAUDE.md based on feedback. Then scale.
10. DAILY LOOP (you, 30 min). Daily-summary script that runs after my last sync each morning, parses sync logs, appends "Sync activity" block to today's daily note. Optional: morning brief at 07:00 reads past 14 days of dailies + recent sessions + sync reports + wiki entities, surfaces patterns, writes today's daily with a "What I noticed" section.
== HOW TO BEHAVE ==
- One question at a time during Stage 1. Don't bundle.
- Pause for confirmation between major steps in Stage 2. Don't blast through.
- If I say "skip", "later", or push back, take my input as priority. Don't argue.
- Defaults from Maran's setup (vault name "Big Brain", launchd 08:00-09:30, etc.) are EXAMPLES, not requirements. Adapt to my Stage 1 answers.
- Plain English. If a sentence needs a glossary, rewrite it.
- Brief progress updates, not narration.
== AFTER WE'RE DONE ==
Tell me to grab the conversation-sync prompt from solstudiosdigital.notion.site/01-obsidian (Downloads section 1) to enable auto-archiving of every Claude Code session into raw/conversations/claude-code/.
Then start using /wiki-query to ask my brain things, /wiki-ingest to add new sources, /save to file the current conversation.
Begin with Stage 1, question 1.