What this tutorial solves
ChatGPT is a great query layer over your notes, but a terrible storage layer. Six months of “chatting with my second brain” leaves people with hundreds of chats they cannot search, export, or trust. PKM with ChatGPT only works when storage stays external (Obsidian, Notion, plain Markdown) and ChatGPT is the lens through which you query and synthesize.
Who this is for
Anyone who has tried “use ChatGPT as my second brain” and discovered three pain points: chats that vanish into the sidebar, no way to link an insight back to the note that produced it, and an embarrassing realization that the AI invented half the connections you cited last week.
When to reach for it
You already have a notes habit — at least 50 markdown files, a Notion workspace you actually open, or a daily journal of more than a month. ChatGPT becomes the retrieval and synthesis layer over that corpus. If you’re starting from zero, build the notes habit first; AI cannot summarize content that does not exist.
When this is NOT the right tool
- People without any notes-writing habit yet — buy Obsidian or Notion first.
- Highly sensitive personal data (mental-health journals, financial detail, medical history) that you do not want sitting on OpenAI’s servers.
- Teams sharing knowledge — ChatGPT Projects are per-user, not collaborative.
- Real-time queries against constantly-changing notes; you only re-upload weekly at best.
Before you start
- Pick your one source of truth and commit. Obsidian + .md is the most portable choice; Notion is fine if you accept its export friction.
- Decide on a snapshot cadence — weekly is the floor, monthly is the realistic upper bound.
- Write a 200-word self-description of the topics, projects, and people in your notes. This goes into Project Instructions so ChatGPT stops asking who Sarah is every chat.
- Set up a “Quotes” folder where you save direct excerpts the AI surfaces. That folder, not the chats, is the durable output.
Step by step
- Pick one source of truth for your notes (Obsidian vault, Notion DB, a folder of .md files). ChatGPT is the lens, not the storage.
- Export a weekly snapshot of your notes as plain Markdown. Name it
snapshot-2026-05-22.mdso date arithmetic is obvious later. - Create a “PKM” Project. Upload the latest snapshot. In Project Instructions, state: “You are a retrieval layer over my notes. Never invent connections. If you cannot find a quote, say so.”
- Use the Project for three things only: retrieval (“Find every note where I mentioned LangGraph”), synthesis (“What have I been thinking about hiring lately, with quotes?”), and connection (“Two notes from different months that contradict each other on the same topic — show both with dates”).
- After each useful session, copy the synthesis back into your real notes app as a new note. Tag the source
#chatgpt-synth-2026-05-22so you can audit later. - Once a month, ask: “What topics am I writing about that have not gone anywhere? Suggest 3 next steps for each.”
First-run exercise
- Take a 1-month slice of notes (not your full archive) and run the workflow on it.
- Ask the same retrieval question two ways — once in plain English, once with the exact tag you used. Compare results.
- Save the first session as your baseline. Note what the model nailed, what it invented, and what it missed entirely.
- Change one variable for the second session: snapshot scope, prompt style, or model. Don’t change all three.
Quality check
- For every “connection” the model claims, demand a direct quote from each side. If it can’t produce both, the connection is fiction.
- Verify dates and note titles — these get hallucinated more than content.
- Ask “what’s missing?” — what topics would a human reader expect that the model didn’t surface? If your snapshot is well-curated, this should be a short list.
How to reuse this workflow
- Save the snapshot script (Obsidian has a “Export to Markdown” plugin; Notion needs a paid export).
- Keep the Project Instructions versioned in your notes app — when behavior drifts, you have a rollback.
- For monthly reviews, reuse the same prompt set. Diffing month-over-month answers is the actual signal.
Recommended workflow
A monthly review: export → upload new snapshot → ask for connections across the month → write 3 short synthesis notes back into Obsidian → ask for next-month focus areas → archive the chat with a clear name like pkm-review-2026-05.md.
Common mistakes
- Treating ChatGPT chats as the place your notes live. Chats are not searchable across accounts, exportable in bulk, or yours in any durable way.
- Uploading sensitive personal notes (mental health journal, finance details) to a default account where chats may be used for training.
- Forgetting to update the snapshot — the Project ends up reasoning over months-old data and confidently fabricating “recent thinking.”
- Letting ChatGPT invent connections that are not actually in your notes. Always demand a direct quote, with note title and date.
- Pinning more than 5 chats. The pin list becomes useless once it grows past a screen.
- Mixing work and personal notes in the same Project. The cross-talk produces weird recommendations.
Advanced tips
- Use file names like
snapshot-2026-05.mdso you can ask “compare snapshot-2026-05 vs snapshot-2026-04” and get a meaningful diff. - For privacy, run a stripped version (no real names, locations, identifying details) through ChatGPT. Use find-and-replace to redact before upload.
- Pair with Obsidian + Smart Connections plugin for local-first AI search; use ChatGPT for the harder synthesis questions Smart Connections can’t handle.
- Add a “do not include” list to Project Instructions: dream journals, draft emails, anything you treat as sketch material.
FAQ
- Is this just RAG?: In spirit, yes — but DIY. A real RAG setup (NotebookLM, custom GPTs with file upload) handles retrieval more automatically and grounds answers to citations.
- NotebookLM vs this?: NotebookLM is better for retrieval and citation-grounded answers. ChatGPT is better for synthesis, generative work, and follow-up dialog. Use both; they’re complementary.
- What about Claude Projects?: Claude handles longer single-document context better. For notes, ChatGPT’s snapshot model and Project UX are still ahead.
- How do I avoid uploading sensitive content?: Keep two vaults: a “shareable” one you export, and a private one that never leaves your disk. Local plugins (Smart Connections, Obsidian Copilot) can search the private vault.
- Why not just paste notes into the chat?: The Project model persists context across chats; pasted text only lives in one chat and disappears from search.