Lessons & Prompt Library

Most people have been using ChatGPT and Claude for over a year. That is hundreds of conversations: research sessions, business strategy, writing drafts, brainstorming, learning. It is all sitting locked inside those platforms doing nothing. Today you get it out.

Path A: Bulk Export (Get Everything at Once)

Path B: Chrome Extensions (Export As You Go)

Use these for ongoing capture after the workshop. Install one now so it is ready.

What Is Worth Exporting

• Business strategy sessions and planning conversations
• Research on topics you care about professionally
• Writing drafts and creative work you want to keep
• Learning sessions on tools, frameworks, or industries
• Key decisions you talked through with AI
• Project scoping or problem-solving conversations

Obsidian vs Notion at a Glance

Notion Workspace Structure

This is the moment where it all becomes real. You have your exports, your vault is set up, and Claude is connected. Now you feed it your history and watch it build.

The Karpathy Flywheel

Every query you run, every output Claude generates: file it back into the wiki. Your explorations become permanent additions to your second brain. Over time, the wiki reflects not just what you have read, but how you think.

The Lex Fridman Method

Lex Fridman feeds his second brain into voice mode and talks to it on long runs. He explains concepts out loud and asks the wiki what he is getting wrong or oversimplifying. This is the Feynman technique (learn by teaching) with a study partner who knows everything in your knowledge base.

Automating the Inbox with Claude Code Routines

The manual approach works. But if you want the system to run itself, Claude Code routines can do it for you.

The key insight: instead of tagging processed files with compiled: true and trusting that nothing gets missed, use physical folder separation. Files that are unread live in one place. Files that have been processed live in another. Claude never has to guess.

Updated Rhythm

• Daily (2 min): Export new conversations → drop in /raw/inbox
• Weekly (5 min): Review what was added to the wiki that week, clean up anything that needs refinement
• Monthly (30 min): Run the Vault Graph Health audit (Module 07)

Closing the Loop: Filing Outputs Back Into the Wiki

Every time you run a query or analysis, Claude produces a response. Most people read it and move on. That response should go back into your second brain.

The logic is straightforward: if the output was worth generating, it is worth keeping. Query outputs represent your thinking at a specific point in time — they capture what your brain noticed, what questions it asked, and what patterns it found. Those are valuable data points, not disposable text.

What Compounds Over Time

• Cross-source queries: save as /wiki/queries/[topic]-[date].md
• Decision reviews: save as /wiki/decisions/review-[date].md
• Career framework output: already saved to /outputs — also link it from /wiki/self/career-analysis.md
• Graph health reports: already saved to /outputs — link from a /wiki page tracking vault health over time

This is what Karpathy means by "the wiki reflects not just what you have read, but how you think." The wiki gets smarter every time you use it — but only if you file the outputs.

The problem with a real second brain is that your knowledge doesn't come from one place. You have Claude exports, ChatGPT exports, meeting transcripts, a Slack workspace history, maybe a Notion database. Each source has its own format, its own naming conventions, and its own way of referring to the same people, projects, and ideas.

If you process them independently, you end up with three pages about the same client, five different spellings of the same person's name, and a wiki that contradicts itself. This module fixes that before it becomes a problem.

The Core Challenge: Entity Drift

The same entity appears differently across sources:

• "Wags" (Slack) vs "Mike Wagner" (meeting note) vs "mike-wagner" (filename) — three references, one person
• "Axe Cap" (internal shorthand) vs "Axe Capital" (formal) vs "the fund" (conversational) — three references, one company
• "the Sandicot play" (chat) vs "Sandicot Municipal Bonds" (document) vs "the trade" (Slack) — three references, one event

When Claude processes each source separately, it creates separate wiki pages for each variation. Your graph fragments instead of compounds.

Source Tagging in Frontmatter

Every page created or updated during multi-source merging should carry a sources field in its frontmatter. This lets you trace any fact back to where it came from and filter your vault by source if you ever need to audit a specific export.

What Success Looks Like

One page per person. One page per project. One page per major decision. Facts that contradict each other flagged and resolved. Source attribution on every page. Your vault stops being an archive and starts being a single coherent record of everything you've built.

Most vaults have a hidden problem: the notes exist, but they don't connect. You can have 1,000 pages and still have 900 of them be islands — no one linking to them, nothing linking out. That's not a knowledge base. That's a filing cabinet.

This module is about graph health: measuring how connected your vault is, fixing what's broken, and building the infrastructure so new notes automatically find their place in the network.

Why Connectivity Matters

Obsidian's graph view is only as useful as the links that feed it. When you ask Claude a cross-source question, it can only synthesize what's connected. Orphaned notes are invisible to that synthesis. The goal is a vault where every note has at least 2–3 incoming links — meaning multiple other notes point to it, and it points back.

What Success Looks Like

After running this module, your vault's average in-degree should be above 1.5, true orphans should be in single digits, and your graph view should show a connected web rather than isolated clusters. Run the audit again monthly as part of your maintenance rhythm.

This module is for people who feel pulled in multiple directions. Maybe you consult and write. Maybe you build products and do strategy. Maybe you collect skills the way others collect opinions. Most career advice tells you to pick a lane. This framework helps you find out which lanes are already feeding each other — using evidence from your own notes, not someone else's theory.

The prompt below runs three analytical frameworks against your vault simultaneously and saves the results as a Markdown report inside your second brain. Run it once your vault has real history in it — at least a few months of notes, projects, and conversations. The output is only as good as what's already there.

Who This Is For

• Freelancers and consultants managing multiple service lines or skill sets
• Career-changers trying to find their throughline across past roles
• Multi-passionate people who aren't sure what to double down on
• Anyone who feels spread thin and wants evidence-based clarity, not gut feel
• Founders and operators whose interests have evolved as their business has

The Three Frameworks

Run the Prompt

Paste this into Claude Code, Claude Desktop, or your Notion MCP session with your vault open. Fill in your vault directory path at the top if Claude cannot find it automatically.

Reading Your Results

• Compounding stacks are interest pairs that scored 3 or higher on the LINK test. These are where you should direct most of your energy.
• Zero-job interests from the Sorting Hat are not necessarily bad — they may be hobbies worth keeping. But if they are consuming professional time without any return, that is worth examining.
• Dormant interests that still hold jobs like Stability or Optionality may be worth reactivating. Dormant interests with zero jobs can probably be officially parked.
• The golden thread in the synthesis section is typically the most surprising output. It names the invisible connector across your work that others already see but you haven't named yet.

You've built the brain, maintained it, merged sources into it, and audited its connections. This module is the payoff: using it as an active collaborator in the work you already do. The prompts here are the ones people bookmark and keep open in a separate tab.

1 — Draft From Your History

Grants. RFPs. Strategy memos. Proposals. Performance reviews. Client briefs. Every document you have ever written is a data point about what works for your specific audience. Most people write each one as if the others don't exist. Your second brain changes that.

The pattern: give Claude the new brief, let it find your most relevant past work, extract what landed and what didn't, then hand you a foundation to build from — not a blank page.

2 — The Digital Twin

Your second brain is not just a record of what you know. It is a record of how you think — your past positions, decisions, frameworks, and the evidence you cited to support them. That makes it possible to do something no other tool can: run a new idea against your own intellectual history.

The digital twin isn't the AI pretending to be you. It's Claude reading your actual notes and arguing back using your own evidence. If you're about to make a decision you made before and regretted, it will find that. If you've shifted on something without noticing, it will name it. If your new idea is genuinely new, it will confirm that too.

3 — Wire It to an Agent

Once your wiki is well-structured and current, it becomes something more than a personal resource — it becomes the knowledge layer for an automated agent. The agent doesn't need to know everything you know. It just needs to be able to read your wiki on demand.

The use case that lands hardest for business owners: an onboarding agent that can answer any question a new team member asks, because every SOP, process, and decision rationale lives in the wiki. No handbook to maintain separately. No Slack threads to search. The agent reads the same source you do.

4 — Private Models

Everything above assumes you're comfortable sending your vault content to a cloud API. For most people, that's fine. But some knowledge is different — proprietary research, client data, medical records, financial models, trade secrets, or anything governed by NDA.

Private models flip the architecture: instead of sending your data to a cloud model, you run the model locally on your own machine. Your second brain never leaves your hardware.

The second brain you've built in this workshop is already structured for this. The /wiki folder, the schema.md, the clean naming conventions — these are exactly what a local RAG setup needs. The day you decide to go private, the vault is ready. The setup is a tooling change, not a content change.

This matters most for: researchers with unpublished data, consultants with client confidentiality obligations, healthcare and legal professionals, and founders who treat their strategy and IP as competitive advantage. The second brain becomes a private intelligence layer that no one else can access or train on.

Use Cases by Persona

One pattern, many applications. Here are six ways real people use a second brain as an active work tool — not just a filing system.

Most second brain disasters are recoverable. Almost all of them are preventable. This module gives you both: the safety net that makes recovery trivial, and the patterns that mean you rarely need it.

1 — Git as Your Undo Button

Every change Claude makes to your vault — every new page, every edit, every deletion — can be tracked as a Git commit. If something goes wrong, you roll back to the last known good state. If something goes partially wrong, you cherry-pick exactly what to restore.

This is the most important five minutes you will spend on your second brain setup. Do it once, forget about it, and let it run silently in the background.

2 — The Safety Prompt Pattern

The best recovery is no recovery. Before any operation that touches a lot of files — a full compile, a lint pass, a multi-source merge, a graph health fix — run a dry-run prompt first. Claude describes exactly what it would do without touching anything. You read it, adjust if needed, then give the green light.

This takes 30 seconds and eliminates 90% of surprises.

3 — Recovery Playbook

Three scenarios, exact steps. All three assume you have Git set up — which is why Section 1 comes first.

4 — When Claude Gets Confused

As your vault grows past a few hundred notes, you will occasionally notice Claude starting to lose coherence mid-operation — creating pages that already exist, referencing files it hasn't read, or producing output that contradicts what you told it earlier. This is a context window problem, not a Claude problem. The vault has simply outgrown what fits in a single session.

The Fix: Scope Down

• Process in batches of 50 files, not the whole vault. Start a fresh Claude session for each batch.
• Run Duplicate Detector first (Module 07) before any large merge or compile. Deduplicating before processing reduces context pressure significantly.
• Use folder-scoped prompts: instead of "process everything in /raw", use "process only /raw/inbox/batch-01". Claude stays focused.
• Commit between batches. Each clean batch gets its own commit so you have clear rollback points.
• Read schema.md at the start of every fresh session. Remind Claude of your conventions before each new batch — it doesn't carry memory between sessions.

The best insight you have all week might happen in the car, on a run, or in the middle of a conversation. If it doesn't get captured in the next 30 seconds, it is probably gone. This module closes that gap — building a path from a phone tap or voice memo directly into your vault without friction, and without waiting until you're back at a desk.

The key connection: everything you capture on mobile eventually lands in /raw/inbox, where the automation you built in Module 05 picks it up and integrates it. You do the capturing. The pipeline does the rest.

1 — Sync Your Vault to Mobile

Before you can capture to your vault from a phone, your vault needs to be on your phone. Three options, honest tradeoffs:

2 — One-Tap Mobile Capture

Once your vault is on your phone, set up a shortcut so creating a new inbox note takes one tap from your home screen — not three taps, a folder navigation, and a filename decision.

3 — Voice to Vault

Voice capture is faster than typing for longer thoughts — a 2-minute voice memo can capture more context than 5 minutes of typing. The challenge is that voice notes are messy: filler words, incomplete sentences, no structure. The prompt in this section handles that.

4 — The Full Loop

When all the pieces are connected, this is what the daily flow looks like for someone using mobile capture:

• Morning walk: voice memo with three ideas → transcribes automatically → one tap to drop transcript into /raw/inbox
• Client call: quick note in Obsidian mobile during the debrief → one tap, 30 seconds
• Evening: inbox automation runs → voice transcripts and quick notes extracted, structured, added to the wiki → files moved to /raw/processed
• Next morning: wiki updated with everything from yesterday without touching a computer

The second brain captures when you're away from the desk and processes while you sleep. The only part that requires you is the initial drop into /raw/inbox — and that takes less than 30 seconds.