Source: raw/I_ve_Set_Up_500+_AI_Second_Brains._Here_s_The_Only_18_Things_You_Need_to_Know.md (primary, newest), raw/Build_a_Better_Claude_Second_Brain_than_99%.md, raw/6_Things_People_Get_Wrong_Setting_up_An_AI_OS_+_Fixes.md, raw/Claude_Cowork_+_Obsidian_Will_Change_How_You_Work_Forever.md
Creator: Ben AI · Platform: YouTube
- Primary: “I’ve Set Up 500+ AI Second Brains. Here’s The Only 18 Things You Need to Know” · URL: https://www.youtube.com/watch?v=pXHpUyC54uQ · Duration: 29:44 · Uploaded: 2026-10-28
- “Build a Better Claude Second Brain than 99%” · URL: https://www.youtube.com/watch?v=Puzgmzv7t3Q · Duration: 18:19 · Uploaded: 2026-09-07
- “6 Things People Get Wrong Setting up An AI OS + Fixes” · URL: https://www.youtube.com/watch?v=lObeYn06BHI · Duration: 23:53 · Uploaded: 2026-06-12
- “Claude Cowork + Obsidian Will Change How You Work Forever” · URL: https://www.youtube.com/watch?v=qo4YZvC1q5I · Duration: 21:54 · Uploaded: 2026-03-24
Ben AI runs an AI agency and an “AI accelerator” community, and says he has helped more than 500 people and businesses set up a second brain. These four videos are one playbook told at four points in time, from his first Cowork + Obsidian setup in March 2026 to an 18-lesson list in October 2026 drawn from his team’s support calls. The thesis does not change: a second brain is a local folder of Markdown files that every AI tool reads from and writes to, routed by one root instructions file. Most of the value comes from five to ten files, and the folder stays useful only if something updates it daily and cleans it weekly. This article records the October list as current and notes where Ben’s advice changed.
Key Takeaways
- A second brain is “a folder on your computer with lots of text files.” Obsidian (or his own app, Baalda) is only a viewer; Claude reads the folder directly. Ben repeats this in all four videos.
- 80% of the value comes from 5 to 10 files. His named set:
me.md,business.md, an ICP doc, a strategy doc,brand.md(voice and style), andteam.mdfor businesses. He says these get pulled into almost every task. - The root instructions file is a map, not a manual. Its main job is a routing table: each folder, what is in it, when to use it, and where to save new things.
- Lock the folder structure. “A fixed folder structure is probably more important than the folder structure itself”, because every change means the routing file has to be relearned.
- Save insights, not raw data. The brain is “a filtered layer on top of” your software, not a copy of it. Connectors still hold the raw data.
- Daily update, weekly cleanup. A scheduled task pulls what changed from your tools every day; an optimizer pass runs every one to two weeks.
- Skills point to the brain, they do not copy it. Keep shared reference files (strategy, voice, offers) in the brain and have each skill reference them, so one edit updates every skill.
- Context compounds, so start now. Ben’s March argument: a competitor who starts six months later is “behind on 6 months of intelligence”, and new models make the same context more useful, not less.
The 18 lessons (October 2026)
Setup (1-7)
- Start with a brain dump. Turn on a dictation tool (he uses Wispr Flow) and talk for about 30 minutes about who you are, the business, customers, goals, writing style and team. Claude turns it into the 5 to 10 core files. His free
/setupskill asks the questions one at a time and writes the files. - Set it up locally. Ben’s reasons against Notion or Google Drive as the primary store: Claude has to go through a connector, which he says is “slower, less accurate, and more expensive”. Built-in Claude memory holds “a few facts” and ties your context to one provider. Hosting trade-offs: Where to Host a Claude Second Brain.
- Visualize it. Obsidian or Baalda opens the same folder and adds a graph view. It “doesn’t really change anything for your AI”.
- Write a good
AGENTS.md. Ben now calls the root fileAGENTS.mdso it works with other providers too. Two rules: a routing table, and under 200 lines, which he attributes to Anthropic. If you already have aCLAUDE.md, rename it and do not keep both, because “when Claude finds a CLAUDE.md in your folder, it only reads that one”. See the Claude Code caveat below. - Use one second brain folder. Businesses, clients and departments are subfolders; the root file routes between them. Exceptions: two businesses with no shared context, or keeping personal and business apart.
- Lock in your folder structure. His example: splitting
clients/into active and past makes Claude keep saving to the old place until the routing file is updated. - Add an index file per subfolder. A small map in each subfolder listing its files and what each covers. Ben credits the idea to Andrej Karpathy: the model “reads the index first to find relevant pages and then drills into them”.
Keeping it working (8-13)
- Save insights, not raw data. His example: AI went through hundreds of support calls, community questions and support tickets to find the top mistakes for this video, and the result is “one simple markdown file”. The brain should also hold what lives only in people’s heads: strategy, decisions, why prices changed.
- Update it every day. A daily scheduled task pulls from your most important connectors and writes only what is new or changed. Meeting transcripts are the most valuable source because “that’s where a lot of decisions get made”. His free
/operatorskill sets this up and also updates or deletes older files the new information makes stale. - Clean it every week. As the brain grows, the root file and index files stop matching the folders, and duplicates, contradictions and orphan notes pile up. His free
/optimizerskill updates the routing and index files, keeps the root file under 200 lines, fixes broken links, merges duplicates, resolves contradictions, trims long files, and ends with a health score out of 100. It fixes small things and asks before big ones. - Build a dashboard on top of it. Three kinds: a general daily overview, one per business area (his sales dashboard means reps “don’t really need to open our CRM anymore”), and one built around a current problem. Skip graph-heavy “agentic OS” visuals that “don’t really give you any value”. Build paths: Agentic OS Command Center; sales version: Claude Sales Second Brain.
- Make skills read from the brain. Ten skills with their own copies of the strategy doc means ten stale copies. Keep the skills themselves in your Claude account, not in the brain (“that gets messy really fast”); only the reference files live in the brain.
- Keep it updating with the laptop closed. Scheduled tasks need the desktop app running. Copy the same prompt into a Claude Code routine (see Routines), which runs on Anthropic’s servers. A routine cannot reach your local folder, so the brain must also be reachable as a connector; Ben uses Baalda’s MCP endpoint for this.
Team rollout (14-18)
- Sync one brain across the team. Ben says Obsidian, Google Drive and Notion setups each fail on at least one of: too technical, edit conflicts, no permissions. Baalda is his answer: invite by email and the vault installs on the member’s machine, or the member only adds the connector (less efficient, he notes).
- Roll out in stages, with one owner. Use it yourself for a couple of weeks, then add one person, then one department, then the company. The owner runs the weekly cleanup and approves changes to the folder structure and root file.
- Control who can see and edit what. Per folder or file: edit, view only, or no access. In Ben’s own brain the context folder and root file are view-only for everyone, and finance is visible only to his co-founder.
- Build team dashboards. Per department, and optionally per person. Share them as a Claude artifact or host them on Vercel.
- Add one “business first” rule to the root file. Ben’s suggested line: “Always act in the best interest of the business based on our goals and our strategy. If a request or output conflicts with that, say so clearly instead of just agreeing.” He cites an a16z piece, “institutional AI versus individual AI”, for the problem: each person’s AI agrees with that person.
The three free skills
Ben ships /setup, /operator and /optimizer as a free plugin ZIP (first link in each video’s description). Install in Cowork via Customize → Plugins → Add → Upload plugin. These are the renamed successors of /os-setup, /os-operator and /os-optimizer in Ben’s Five-Skill AIOS Setup.
| Skill | What it does | Ben’s model advice |
|---|---|---|
/setup | Asks solo or business, builds the folder structure and root file, then runs a 12-topic interview (you, company, market, offer and so on). Each topic has a free-text box, a box naming software to pull from, and file upload | Not Fable 5.1: “unnecessary token spend”. Opus 5 is enough |
/operator | Asks how often to run (at least daily) and which sources to pull from (he recommends the meeting recorder, email or Slack, and the CRM). Creates the scheduled task and saves its instructions as a prompt file in the brain | Not stated |
/optimizer | Audits the whole brain against “Anthropic, Andrej Karpathy, and many others” (the September video’s list includes Karpathy’s wiki guide and “the caveman method”), applies fixes, outputs a dashboard report | Fable 5.1: “extremely good at spotting these inefficiencies in large sets of context” |
Optimizer runs Ben has shown (his own vault, his numbers):
| Video | Files audited | Issues found | Fixes applied | Model |
|---|---|---|---|---|
| May 2026 (five-skill video) | 1,700 | 34 | 32 (health 46 to 94) | not stated |
| June 2026 (6 Things) | 1,700 | not stated | 32 | not stated |
| September 2026 (Better than 99%) | 5,400 across 822 folders | 56 | 26 | Fable 5.1 |
The May and June runs look like the same run shown twice.
Folder structure
Ben’s “80/20” layout, from the June video. The /setup skill builds it.
| Folder | Holds |
|---|---|
context/ | The 5 to 10 core files: me, business, ICP, strategy, brand, team |
daily/ | One file per day, written by the operator task: meetings pulled, escalations, calendar, Slack and community activity |
intelligence/ | Meeting transcripts, decision logs, market and competitor research, reports |
projects/ | One subfolder per project (per client for an agency; per YouTube video for Ben) |
resources/ | Reusable templates, prompts, SOPs, old posts |
team/ | Businesses only: a profile per person with their own daily file |
The March business version also had departments/, onboarding/, skills/ (skill reference files) and tasks/. By June Ben had cut it back: “less is more”, and add a folder only when you keep using Claude for something new.
Where Ben’s advice changed
- Root file length: “under 300 lines” in June; “under 200 lines” in October, now attributed to Anthropic. Anthropic’s own figure is 200 (see Steering Claude Code).
- Root file name:
CLAUDE.mdin March to September;AGENTS.mdin October. - Skill reference files: in March he kept them in a
skills/folder in the brain; in October he says keep skills in the Claude account and point them at brain files. Both versions keep the reference files in the brain. - Team sync: Obsidian Relay plus his own permissions fork until May; his Baalda app from late August. See hosting options.
Claude Code caveat on AGENTS.md
Ben’s rename advice matches how Claude Code reads the files: since v2.1.277 (2026-09-18) Claude Code reads AGENTS.md only when no CLAUDE.md or CLAUDE.local.md sits at or above the working directory (see AGENTS.md). Two things he does not mention: Claude Code builds before v2.1.277 ignore AGENTS.md entirely, and the fallback was not yet on Bedrock, Vertex or Foundry. Keeping a one-line CLAUDE.md that imports @AGENTS.md is an alternative to renaming.
Try It
- Create an empty folder, select it in Cowork or Claude Code, and dictate a 30-minute brain dump. Ask Claude to turn it into
me.md,business.md, an ICP doc, a strategy doc andbrand.mdin acontext/folder. - Ask Claude to write a root
AGENTS.md(orCLAUDE.md) under 200 lines whose main section is a routing table: folder, contents, when to read, where to save. - Create the six folders above and do not rename them for at least a month.
- Add an index file to any folder with more than about 20 files.
- Set up one daily scheduled task that reads your meeting recorder and email and writes a
daily/YYYY-MM-DD.mdwith only what changed. - Run a cleanup prompt (or Ben’s
/optimizer) every one to two weeks: update the routing and index files, merge duplicates, flag contradictions, fix broken links. - Edit one existing skill so its reference files point into the brain instead of living inside the skill.
- For a team: add the “business first” line to the root file, name one owner, and roll out to one person before the whole team.
Open Questions
- The a16z “institutional AI versus individual AI” article is cited but not linked in the transcript; not located.
- Baalda “is now an official connector inside of Claude” (Ben, October): not verified in Anthropic’s connector directory.
- “500+ setups” and “hundreds of support calls” are Ben’s own figures for his paid community and app users; unverified.
- Which frameworks the optimizer combines. Ben names Anthropic, Karpathy and “the caveman method” and says there are about nine; the skill files themselves have not been read for this wiki.
- Does Cowork’s folder instructions panel read
AGENTS.md? The Claude Code behavior is documented; the Cowork behavior was not checked.
Related
- Ben’s Five-Skill AIOS Setup — the May 2026 version of the same skill kit
- Where to Host a Claude Second Brain (Ben AI) — local, skill, Drive, Notion, Obsidian + Relay, Baalda
- Agentic OS Command Center (Ben AI) — dashboards on top of the brain
- Claude Sales Second Brain (Ben AI) — the same pattern for a sales team
- Every Level of a Claude Second Brain — Nate Herk’s retrieval levels and Ben’s seven context levels
- Five AI OS Organization Hacks (Nate Herk) — a parallel set of folder and routing rules
- AGENTS.md — how Claude Code reads AGENTS.md versus CLAUDE.md
- The 2026 Claude Code AIOS Pattern — where Ben’s advice sits among other practitioners