I had a wonderful weekend building my second brain — filling gaps in how I think, and making sure Claude works with me the way I want it to.
These are concept pictures, not blueprints. The flow is very opinionated — so there's no point sharing the wiring. The reason for sharing is: you can do this too.
1 — The Operating System
Capture · Distill · Decide · Review.
Inputs from life — emails, books, meetings, web articles, ideas, conversations — pour into a single funnel. Distill scripts (powered by Claude Sonnet) turn raw into shaped. Three filters decide what survives: role coverage, area-of-responsibility scope, time + energy. Action splits clean — do, delegate, wait, kill, reference, trash. Review runs on a cadence: daily, weekly, monthly, quarterly.
The whole thing rests on Four Pillars + Rigor.
2 — Four Pillars + Rigor
The principles under everything.
Truth — verified evidence over plausible guess. Tool every claim this session.
Effort — sharpen the axe (40:10). Forty minutes reading, ten cutting.
Efficiency — DRY · right tool · agent over main context · parallel where independent.
Effectiveness — solve the user's actual problem. Verify outcome, not shape.
Rigor is the method that delivers all four: Sharpen · Map · Mechanism · Snapshot · Verify · Guard.
Anti-DNA — the patterns to break: Patch · Bespoke-design · Honor-system · Phantom-gap.
3 — Open Brain + Wiki: How a Question Gets Answered
The brain is a Postgres + pgvector library, fed continuously from two Obsidian vaults.
When I ask Claude a question, a protocol forces a search of the brain first — never memory. Hybrid search (semantic + keyword + recency) returns chunks with their original sources. Each claim in Claude's answer carries a [chunk:N] citation. If zero results: Claude says so explicitly — "I searched and got nothing; what I'm about to say is from this conversation, not your brain."
Memory files are not ground truth. Citations are.
4 — How Claude Stays Aligned (Every Edit, Every Session)
Every Edit or Write fires a hook. Multi-trigger by design:
First-prime per session — full foundations primer surfaces on the first edit
Context-shift — when I cross a project boundary mid-session, re-anchor
Anti-DNA detector — on the 3rd edit of the same file, surface "patch detected — restructure the substrate, not write patch #4."
Whitelist includes Markdown, so writing also gets primed. State is per-session; /clear resets cleanly.
Declaration is not a mechanism. A trigger that fires IS.
5 — The Safety Nets: G-guard Chain
Canon → Surface → Hook → Drift detectors.
Five canon files live as scrolls. They're distilled into one runtime reminder. The hook reads it. And four scheduled night-watchmen patrol the vault + codebase:
pillar-lint— Sundays 19:30, audits 149 filesopenbrain-contradictions-lint— daily 03:45staleness-detector— daily 03:50 (new this weekend)kamakshi-lint— Sundays 19:00
You can't depend on good intentions, only on good mechanisms.
A grep that runs in lint IS a mechanism.
Built with Claude Code over a weekend, on top of Open Brain (self-hosted on a small VPS) and two Obsidian vaults. The point isn't the wiring — it's that you can build your own.
Credits — ideas I borrowed from
This weekend's work stood on a lot of other people's shoulders. Crediting the ones whose thinking shaped specific pieces:
Andrej Karpathy — for the "personal wiki" frame that started me thinking about second brains as text-first systems.
Tiago Forte — for PARA (Projects · Areas · Resources · Archives), the folder bones under both my Obsidian vaults.
David Allen — for Horizons of Focus (Roles → AORs → Projects → Top 3 → Ground), which still anchors my Operating System chart.
Stephen Covey — for "roles before goals", which became the
role/a..htag spine in my vault.Steph "kepano" Ango — for
github.com/kepano/obsidian-skillsand showing what well-shaped Obsidian skills look like. I now use hisobsidian-cli,obsidian-bases, andjson-canvasskills directly.Mansel Scheffel — for the 3-layer skill chaining pattern (context fork → file handoff via
temp/→ bang-command injection). Cut his lead-research skill from 51K → 5–8K tokens per run; I'm now applying the same shape to my daily skills.Peter Pang (CreaoAI) — for the self-healing agent harness (tri-judge eval + 6-job pipeline + AI-gated grey rollout). I haven't built it yet — it's the next stream of work.
Ole Lehman — for The Council (5-advisor peer-review skill: Contrarian · First Principles · Expansionist · Outsider · Executor). I rebuilt a version from spec; my
/councilskill is downstream of his thinking.Charlie Hills — for surfacing The Council in Issue 042 of his Substack newsletter, where I first ran into it.
Nate B Jones — for the OpenBrain thesis (hybrid wiki + DB + citation discipline) that Open Brain is built on.
Zaid (
Kajin-0/Mercury) — for the contradictions-lint and staleness-detector failure modes I engineered against this weekend.Manthan Patel — for the ".claude/ folder, fully mapped" post on LinkedIn that pushed me to bootstrap canonical layouts across all my projects.
Anthropic — for Claude Code itself (and the SDK / Agent Skills spec the rest builds on).
My partner and family — for time stolen from what was rightfully their time over the weekend.
