Every AI conversation starts from zero. The model doesn't know your stack, your clients, your taste, or your past decisions, so you re-explain everything every time — a tax that compounds until you just settle for shallow answers. A second brain solves it. Not a Notion vault, not a folder of notes — a small, curated set of markdown files designed to be read by an LLM, organized around what the AI needs to know to be useful instead of what feels productive to organize. Three files carry 80% of the value: an identity file (who you are, how you work, what you refuse), an active projects file (one paragraph each, updated weekly), and a decisions file (every meaningful choice and why). My biggest mistake was making it comprehensive. I dumped everything in and the AI got worse. The fix was the opposite of what felt productive: delete most of it. A second brain is curation, not collection. The compounding effect — by month two the AI sounds like a coworker who remembers — is real. The discipline required to update three files weekly is the bar most people fail to clear.
Every conversation with an AI starts from nothing. It doesn't know you switched stacks last month. It doesn't know which client hates Tailwind. It doesn't know your tone of voice or that you decided three weeks ago that you'd stop suggesting Docker for anything below twenty containers. So you re-explain. Every. Single. Time.
That re-explanation is a tax. The first time you pay it, it feels normal. The hundredth time you pay it, you stop paying it — you just accept shallow answers because giving full context costs more than the answer is worth. You ask "what should I do here?" and you get the median internet response, because that's all the model has to work with.
A second brain is the fix. The model still starts blank, but you hand it a curated slice of context the moment it needs it. Done well, this turns a generic chat bot into something that feels like a coworker who's been with you for six months. Done badly, it's a folder of stale notes that makes the AI worse instead of better.
It's not a folder of notes. It's not a Notion workspace. It's not a graph of bidirectional links. It's a structured set of markdown files designed to be read by an LLM, organized around what the AI needs to know to be useful — not around what feels productive to organize.
The distinction matters. Personal knowledge management for humans is a different problem from context engineering for AI. Your beautiful Roam graph with twelve hundred connected notes is, to an AI, an exhausting source of noise. The reverse is also true: a thin file structured for AI consumption is, to a human, dry and lacking the connective tissue we use to find things.
The setup I use:
The whole thing is under 50 files. Most of them are under 200 words. The constraint is not technological — Claude can read 200,000 tokens at a time — it's editorial. Less context, better answers.
Three files do roughly 80% of the work. If you only ever build these three, the impact compounds anyway.
1. Identity. Who I am, what I do, how I communicate, what I refuse to do. My role and primary projects. My tone preferences ("direct, no fluff, push back when warranted"). The languages I work in and when to use which. The decisions I've made about my own life that aren't open to debate. The AI reads this once at the start of any session and stops giving generic life advice or generic technical advice. Length: about 400 words.
2. Active projects. A one-paragraph summary per project I'm currently working on. Status, stack, current blockers, who's involved, key recent decisions. Updated weekly without exception. The AI uses this to figure out which project a question relates to and to bring relevant context to bear without me having to spell it out. Length: about 800 words, depending on how many projects are active.
3. Decisions. The most underrated file in any AI workflow. Every meaningful technical or operational decision I've made and the reason I made it. "We're using Vite, not Next.js, because most client work is SPAs." "We're hosting on Vercel, not Lovable, because we needed apex domain support." "We don't take projects under $3K, because the overhead doesn't justify the rate." The AI stops re-suggesting options I already rejected, which is the single biggest source of friction in long-term AI collaboration. Length: about 1,000 words and growing slowly.
Everything else is optional. Glossaries, external references, meeting notes, vendor logins — they all have their place, but they're additive. Most people over-engineer this part, build elaborate vault structures, and then never maintain anything. Start with three files, not thirty.
The first version of my second brain was comprehensive. I dumped every note, every PDF, every Slack export, every old email I could find into one vault, thinking more context would produce smarter answers. I'd watched the demos. I'd read the threads. Surely the model gets better with more to read.
The AI got worse, not better. It kept surfacing irrelevant context. It contradicted itself within the same conversation. It would cite a note I'd written eighteen months ago that no longer reflected my opinion. It would optimize for the wrong constraints because it found an older constraint I'd long since dropped. It was reading too much, drawing the wrong conclusions, and presenting them with confidence.
The fix was the opposite of what felt productive: I deleted most of it. I cut the vault from ~400 files down to ~30. I archived anything older than six months unless it was load-bearing. I rewrote my three core files as if I were writing them for the first time, with the benefit of having seen what the AI actually consumed.
The AI immediately got better. Faster. Less hedged. More opinionated in ways that matched my actual opinions instead of averaging across stale ones.
A second brain is curation, not collection. The hard work is deciding what NOT to include. Every file you keep should earn its place, and "I wrote this once and might need it" is not a good enough reason.
Two modes, depending on the tool:
Both work. The point is the model never starts cold for anything that matters. The first message of any new session is already grounded.
I also keep a memory file specifically for the agent to write to — a place where it can record things it learns about me, projects, decisions, or preferences during a session, so the next session inherits that knowledge. The agent maintains it itself, with rules for what gets recorded.
The first week, this feels like overhead. You're writing files for an AI to read, the AI hasn't yet learned to use them well, the answers don't immediately feel different. The temptation is to abandon the experiment.
Push through the first week. By the second month, you notice the AI is making suggestions that sound like you — referencing past decisions, matching your tone, refusing the approaches you've already ruled out, pulling up project context without being prompted. You stop saying "remember that the stack is X" because you no longer have to.
That's the moment it stops being a chat bot and starts being a coworker who remembers things. The shift is felt, not measured. Once you've felt it, you don't go back.
The discipline most people fail to develop: updating the three core files on a fixed schedule.
I update active projects every Friday afternoon as part of my weekly review. It takes ten minutes. I update decisions whenever I make one — not in batch, but immediately. I update identity when something meaningful changes about how I work or what I'm prepared to do.
If you don't maintain it, the second brain decays into a museum of how you used to think. That's worse than not having one at all, because the AI will confidently use stale context.
A useful rule: if you stop updating for a month, archive everything and rebuild with the three core files. Don't try to patch a stale vault back into accuracy.
Two things to think about before you start:
Privacy. Your second brain will contain client names, salary numbers, sensitive decisions, and opinions you would not want public. Keep it local (Obsidian stores files on disk), or in a personal git repository you control, or in a Claude Project you don't share. Do not upload it to a vendor that uses your data for training.
Portability. Use plain markdown, not a proprietary format. The whole point of this exercise is to own the artifact. If you store it in a tool that locks you in, you've just moved the lock-in problem from your AI vendor to your note-taking app.
The compounding works. But only if you keep paying the small upfront cost of curation. That's the price of admission, and it's much smaller than the tax you pay re-explaining yourself for the rest of your career.