An AI OS (or "second brain") is a structured folder of context files that an agent like Claude reads to do your work with your real context instead of generic guesses. Most setups fail the same way: too much raw data, no structure, a bloated memory file, and inconsistent use. In a 24-minute video, Ben AI pulled the six most common mistakes from hundreds of setup support calls, with the fix for each. This note captures all six, splits the takeaways by audience, and ends with a prompt for fix #5, the second-brain audit.
The six mistakes and fixes
| # | Mistake | Fix |
|---|---|---|
| 1 | Wrong or thin context (raw data dumps) | Build the 5-6 essential context files that carry 80% of the value. Answer the core who / what / why instead of dumping everything. |
| 2 | Context goes stale | Keep a dated daily log and pull real-time context (meetings, calendar, tasks, messages) on a schedule. |
| 3 | No fixed folder structure, or one too complex | Pick one 80/20 structure and stick to it. Consistency matters more than the exact layout, because the agent learns where to read and write. |
| 4 | Bloated CLAUDE.md | Keep it under 300 lines, add a routing table, and use a small index CLAUDE.md per subfolder. |
| 5 | Never auditing the second brain | Regularly dedupe, resolve conflicts, delete blank files, fix broken links. Run an optimizer weekly or biweekly. |
| 6 | Not using it consistently (the biggest one) | Make the AI your default for real work. Context and skill both compound only with use. |
The thread through all six: less is more. Start simple, put the effort into context, and be consistent.
For business people
- Why it pays off. An AI OS compounds. The value is your unique context plus your skill at directing AI, and both grow only when you actually use the system. In a few months the gap is large versus someone who never adopted it.
- What to prioritise. Context first (mistake 1), then consistency (mistake 6). Those two carry most of the result. The folder and file mechanics matter, but they are downstream of having good context and actually using it.
- A new model will not rescue a weak setup. Your context is the moat, not the model. Ben AI argues an older model with good context already beats the newest model without it, and context windows are not expected to jump soon.
- Teams: make one person accountable for maintaining the second brain, even with automation in place.
For technical people
The concrete setup behind the fixes.
-
Folder structure (the 80/20 layout):
context/the essential, always-relevant filesdaily/dated logs of what happened each dayintelligence/meeting transcripts, decisions, market researchprojects/one subfolder per project or clientresources/reusable templates, prompts, SOPsteam/per-person profiles and notes (teams only)
Add a subfolder only when a real recurring need shows up.
- CLAUDE.md is the map the agent reads every time. Keep it under 300 lines; include a routing table that says what context lives where and where to save new context; add a short index CLAUDE.md inside each subfolder. Follow the Anthropic and Karpathy best practices rather than letting the model write a bloated one.
- Maintenance. As it grows you accumulate duplicates, conflicts, misfiled and outdated notes. Audit on a schedule: dedupe, resolve conflicts, delete blanks, fix broken wiki links, and reconcile the routing table against the actual folder tree. A saved routine or scheduled task does this without manual effort.
- Model choice. Pair a capable model with good context. Context beats chasing the newest release.
A prompt to audit your second brain (fix #5)
This prompt does only fix #5: the regular audit and clean-up of the second brain, mirroring the operations Ben AI's OS optimizer runs on the slide. Point an agent with file access to your second brain (Claude Code, or Claude Desktop with filesystem access) and run it:
You are auditing and cleaning my AI OS (a second brain at [PATH], used with Claude).
Do only the audit below, nothing else. First show me what you found; then apply the
safe fixes and end with a dashboard-style report: number of files audited, fixes
applied, and a per-item changelog. Never delete anything without listing it first.
Run these checks across the whole second brain:
1. Duplicates. Find duplicate and near-duplicate notes. Propose one merged version
for each set; keep it and list the others to remove.
2. Routing vs reality. Compare the routing table in CLAUDE.md against the actual
folder structure. Flag every mismatch and update the routing table to match.
3. Blank files. Find empty or near-empty files and remove them (list them first).
4. Conflicting or outdated info. Find notes that contradict each other or hold stale
facts. Surface each conflict and say which version looks current.
5. Broken links. Find broken [[wiki links]] and fix or flag each one.
6. Formatting. Normalise inconsistent formatting (headings, front matter, tags) so
notes follow the same structure.
End with the report plus a short list of anything risky you left for me to decide.
Run it on a weekly or biweekly schedule (or as a scheduled task) so the second brain stays clean as it grows.
Key takeaways
- Context is the moat. The biggest wins are the essential context files and keeping them current, not the model.
- Consistency beats cleverness. A simple structure used every day outperforms an elaborate one you abandon.
- CLAUDE.md is leverage or drag. Under 300 lines with a routing table and per-folder index files; audit it regularly.
- Less is more. Start simple, add structure only when a real need appears.