Working with Claude Resource

An AI second brain — choosing your level

A keep-it-by-your-side sheet for building a second brain your AI can use. For each of the five levels you get the question that selects it, the move to make, and the wall that tells you it's time to climb. Open it when you want to find out which level you're actually on — and whether it's worth going higher at all.

This is the version to keep open while you organize data for your agent. Work in order: set the rule first, then diagnose your level by the pain you actually feel. Don't start with the technology — start with the question your second brain has to answer.

The guiding rule in one line: going up is not better — you're looking for the lowest level that removes your pain.

Before you choose — one rule

Three things to settle before you move anything up a level.

  • Work backwards. Start from "how will I want to ask for this later?" How you'll retrieve the data decides how you should store it now — like a ball and a hoop: the hoop is round, so you'd never make the ball a cube.
  • Measure one thing: will you find it. The real test of a second brain isn't "how much did you put in," it's "will you and your agent find this in six months." If the answer is no, the problem is how the data is organized, not the model.
  • One project doesn't live on one level. Different folders can sit at different levels. Diagnose each on its own, and don't climb where nothing hurts.

Level 1 — find it by exact word or file

  • The question: can I find a file or fact by an exact word or name?
  • What it gives you: one routing file — CLAUDE.md in Claude Code (AGENTS.md in other tools) — that loads automatically and tells the agent who you are, how you work, and where things live. The rest is plain markdown folders: context, a decision log, projects.
  • Your move: make one folder, drop your notes in, add a short CLAUDE.md about you and your setup — with routing rules ("information about me lives in this folder").
  • The wall: the routing file grows, gets messy, and starts to feel ignored. That's your signal to move to level 2.
  • You're here if: you're still re-explaining your setup and searching by exact word or file name. Most people can stop here.

Level 2 — pull everything on one topic together

  • The question: can I pull together everything I know on one topic?
  • What it gives you: a knowledge wiki — files with indexes the agent drills down through. It starts at the topic's index page, reads it in full, follows the trail to the next page. This is also where auto-memory earns its place (in Claude Code you switch it on with /memory).
  • Your move: ask the agent to write an index page and a one-line summary per note, then cross-link them.
  • The wall: the wiki grows until reading whole pages gets slow or imprecise — or you keep searching with different words than you wrote.
  • The trap: "see also" links look like a knowledge graph but aren't — they lack named relationships (that comes at level 4). A visual tool like Obsidian is just a view of the same markdown files; it adds no capability your agent doesn't already have.

Level 3 — search by meaning

  • The question: what if I search with different words than I wrote?
  • What it gives you: semantic search backed by a vector database — you chunk a document, turn each chunk into a vector (an address in "meaning space"), and the system returns the nearest by meaning. Search "feedback" and it also surfaces "test results" and "evaluations."
  • Your move: move your notes into a tool with semantic search and let the agent help configure it. Keep it alongside keyword search — you mix, you don't replace.
  • The wall: chunking loses full context. For a question like "summarize the March 5th meeting" or "which week had the highest sales?", the system sees only a few chunks and can miss the whole.
  • You're here if: you keep whiffing on notes you know exist. For most creators this is the top worth reaching for.

Level 4 — relationship chains (a knowledge graph)

  • The question: can I trace relationships — ask about topic X and follow it back to topic A?
  • What it gives you: a knowledge graph — entities (people, companies, projects) plus named relationships between them ("Acme is endorsed by Postpilot; Postpilot is a competitor of Cadently"). Light to read: the agent reads just the related node, not the whole page.
  • Your move: gather enough material first — use the interview method (let the agent interview you question by question; feed it transcripts and notes). Without data, a graph has nothing to build relationships from.
  • The wall / when to skip: this is usually the skip rung. Your real questions are rarely relationship chains, and a graph is the most complex level and often the most expensive — especially on a managed platform.
  • A privacy note: sending all of this to a cloud model means that data is no longer private. For sensitive client data, consider local or open-source models.

Level 5 — an always-on, autonomous brain

  • The question: what if I want all of this to happen on its own?
  • What it gives you: everything from the lower levels plus an always-on quality — a brain that syncs, refreshes its memory, and adds knowledge by itself. You can assemble it in Claude Code, but you'll manage the scheduled jobs (the crons) yourself.
  • The wall / risk: too much context does more harm than good. When the brain ingests everything automatically, it's easy to clog it with noise and make answers worse.
  • You're here if: you run agents offline and sync knowledge across many agents. This is a level for very specific needs, not "the best" one.

Diagnose your level

You don't pick a level by ambition — you diagnose it by pain. Go folder by folder:

  • Re-explaining your setup and searching by exact word or file name → level 1.
  • 30+ notes and you keep forgetting what's in them → level 2.
  • Routing keeps whiffing on notes you know exist → level 3.
  • You need relationships and chains of thought — following topic X back to A → level 4.
  • Running agents offline and syncing knowledge across many agents → level 5.

What to feed it — context versus connections

What you let in matters more than which level you pick.

  • Context — keep it in. Evergreen, holistic data: what the business has done, locked-in decisions, this quarter's goals. Ask of each item, "in a year, will this memory still help me?" If yes, it belongs.
  • Connections — keep them out. Volatile data: Slack threads, emails, live customer records. Ingest it permanently and it becomes noise you'll clean out every month.
  • But "out" doesn't mean "cut off." Give your brain access to fetch the volatile source live when it's genuinely needed. That creates a retrieval cascade: the goals file first, then the wiki and transcripts, and only as a last resort the live source.

The smallest move today: take one project you already have and ask yourself one question — if I came looking for something specific here in six months, would I find it? The answer shows you which level you're really on.