Obsidian Second Brain Template: The 3-File System I Actually Use

TL;DR

Most "Obsidian second brain templates" you find online are designed for human note-taking — daily notes, MOCs, zettelkasten links, dashboards. They're beautiful and they don't help your AI at all. The template below is the opposite: three short files, written for an LLM to read, that turn Claude or ChatGPT from a generic chat bot into something that sounds like a coworker who remembers. Identity.md, projects.md, decisions.md. Total length under 50 lines each. Updated in 15 minutes a week. Fork it, fill in your own answers, and load it into any AI session. The compounding effect — by month two the AI starts citing your past decisions — is real. The mistake to avoid is treating this template like a normal Obsidian vault. It isn't. It's a context payload for an AI.

Why Most Obsidian Templates Fail for AI

Search "Obsidian second brain template" and you'll find vaults with daily notes, MOCs (maps of content), zettelkasten links, PARA folder hierarchies, and dashboards with 14 plugins. They are designed for a human to read, retrieve, and re-encounter ideas over time.

When you hand one of those vaults to Claude or ChatGPT, the AI does poorly. Not because the model is dumb — because the vault is structured for the wrong reader. Daily notes are noisy. MOCs are recursive. Backlinks don't matter when the AI can read every linked file anyway. Half the content is meta-organization that adds tokens without adding signal.

A second brain template for AI throws all that out. What it keeps is what the AI actually uses to be useful: who you are, what you're working on, and what you've already decided. Everything else is optional.

The Three Files

Open Obsidian. Create a new vault called \brain\ (or whatever you like). At the root, create three files:

1. identity.md

Who you are, how you work, what you refuse to do. The AI reads this first and stops giving generic advice.

\\\`markdown

Identity

Role

[Your job/role in one sentence. E.g. "Senior frontend engineer at a startup,

freelance on the side."]

What I do

[Three bullets max — the actual work you spend time on.]

Communication

  • Direct, no fluff
  • Push back when warranted
  • [Anything else: language preferences, tone, jargon allowed]
  • Languages

    [Which languages you work in. When to switch. E.g. "English by default,

    Arabic in casual messages."]

    What I refuse

  • [Hard nos. E.g. "Not interested in crypto projects."]
  • [E.g. "Won't suggest Docker for fewer than 20 containers."]
  • What I'm optimizing for

    [One sentence on the larger goal. E.g. "Shipping fast, low maintenance,

    hand-off ready."]

    \\\`

    That's about 300–400 words once filled in. Resist the urge to write more. Resist the urge to add subsections for hobbies, family, or hopes and dreams unless they shape your work.

    2. projects.md

    A one-paragraph summary per active project. Updated every Friday afternoon as part of your weekly review.

    \\\`markdown

    Active Projects

    [Project A]

    Status: [In progress / blocked / shipped]

    Stack: [3–5 line stack summary]

    Recent decisions: [What changed in the last 2 weeks?]

    Blockers: [What's blocked?]

    Key people: [Who else is involved?]

    [Project B]

    ...

    \\\`

    About 600–1000 words total once you have 3–5 active projects. Anything that hasn't been touched in 30 days moves to \archive/projects-archive.md\.

    3. decisions.md

    The most underrated file in any AI workflow. Every meaningful decision and the reason for it. Append-only.

    \\\`markdown

    Decisions

    Stack

  • Use Vite, not Next.js, for SPAs — most clients don't need SSR. (2026-03)
  • Host on Vercel, not Lovable — needed apex domain support. (2026-04)
  • Use Supabase for backend — Postgres + auth + storage in one. (2025-12)
  • Business

  • Don't take projects under $3K — overhead doesn't justify rate. (2026-02)
  • Quote in USD, invoice in TRY for Turkish clients. (2026-01)
  • Tools

  • Claude > ChatGPT for code. (Confirmed 2026-05)
  • Obsidian > Notion for personal notes — local-first matters. (2025-09)
  • \\\`

    This file grows slowly — maybe 10–20 entries a month. Each entry is 1–2 lines. The format that works: \decision · reason · date\. Don't over-format.

    Loading the Template

    Two ways to use it depending on the tool:

    Claude Code: Save the three files at the root of your project directory. Rename \identity.md\ to \CLAUDE.md\ and the project picks it up automatically on every session. For per-project context, add a project-specific \CLAUDE.md\ inside the project folder — it stacks on top of the root one.

    Claude.ai or ChatGPT web: Keep a "starter pack" prompt saved as a snippet. Paste the contents of all three files at the start of any meaningful session. Clunky compared to Claude Code but covers the chat use case.

    Both routes give you the same outcome: the AI never starts cold for anything that matters.

    The Mistake to Avoid

    The biggest mistake people make with this template is treating it like a normal Obsidian vault — adding daily notes, building MOCs, linking everything to everything else.

    Don't. The vault is a context payload, not a knowledge base. Every file you add competes for the AI's attention. Every backlink the AI doesn't understand is noise. Every "meta" file (templates, dashboards, indexes) is tokens spent on infrastructure instead of substance.

    If you want a traditional Obsidian PKM setup and an AI second brain, keep them in separate vaults. They have incompatible design goals.

    The Maintenance Habit

    The discipline that separates a useful second brain from a stale one: update \projects.md\ every Friday afternoon. Non-negotiable.

    It takes ten minutes. Open the file, walk through each project, ask:

  • Is the status still accurate?
  • What's blocked that wasn't blocked last week?
  • What decisions did I make this week that should move to \decisions.md\?
  • Has anything been dead for 30+ days? Archive it.
  • Friday afternoon is intentional. Mondays are reactive; you don't have perspective. Fridays you do. Put it on your calendar. If you skip three Fridays in a row, your vault is stale, and a stale vault makes the AI worse, not better.

    What the Compounding Looks Like

    Week 1 feels like overhead. You wrote files for an AI and the answers don't feel different.

    Week 4, you notice the AI stops re-suggesting Docker. It starts asking about your existing stack before recommending changes. It stops adding "depending on your use case" qualifiers because it knows your use case.

    Week 8, it sounds like you. References past decisions. Matches your tone. Refuses approaches you've already ruled out. You stop writing "remember that the stack is X" at the top of every message because you no longer have to.

    That's the shift. It is felt, not measured. Once you've felt it, you don't go back.

    What's Not in This Template

    Things people ask about that aren't here, and why:

  • Daily notes — useful for humans, noise for AIs. Keep them in a separate vault if you want them.
  • Meeting notes — append the decisions from a meeting to \decisions.md\. Skip the rest.
  • Reading notes / highlights — only the ones that changed your thinking. Add them to \decisions.md\ as "I changed my mind on X because of Y (source)."
  • Client portfolios / case studies — public-facing content, not AI context. Keep elsewhere.
  • Glossaries — only if you work across teams that use the same term differently. Most people don't need this.
  • Start with three files. Add only when you've felt the absence in a real session.

    Forking the Template

    The template is intentionally simple enough that you don't need to download anything — three files, copy the markdown above, fill in your own answers. No plugins, no Templater syntax, no Dataview queries.

    If you want a starting point, the pillar second-brain article walks through the philosophy in more depth, and the Steph Ango skills setup covers how to layer reusable AI workflows on top of this base.

    The whole point is that a second brain isn't a tool problem. It's a curation problem. The template is just the cheapest possible scaffold to get out of your way.