Second-Brain OS wires Claude Code into your files, your calendar, your inbox, and a structured second brain that remembers what you're working on. The tools install themselves. Your job is to show up.
We build harnesses because managing context is the hard part of working with AI. Every prompt to a vanilla chatbot is a cold start. The Second-Brain OS gives Claude persistent memory of you, your tools, your work — so every conversation picks up where the last left off.
You already know what an OS is — macOS, iOS, Windows. The layer between you and your computer where everything lives. You don't think about it because it just works.
Now add AI on top. Agents. Persistent memory. Reasoning that runs while you sleep. The OS doesn't just hold your files — it can see them, connect them, and act on them. It can read your inbox, scan your calendar, draft your reply, remember what you said in a meeting three weeks ago, and tell you what matters this morning.
That's Second-Brain OS — your AI Operating System. It's what this onboarding builds.
Build in this order. You can't have Cadence without Connections. You can't have Capabilities without Context.
SOUL.md, USER.md, project memory.gws, gh, databricks), APIs.Claude Code today. Codex tomorrow. Something else next year. The tools churn — that's fine.
What survives is the harness: the folder structure, the CLAUDE.md + memory.md discipline, the skills you wrote, the context you accumulated. Port that to whatever agent comes next and you're moved-in within an hour. That's the real asset you're building. The agent runtime is just the host.
You're not learning a tool. You're building a layer of durable, agent-readable context about your work. The tool is replaceable. The layer is not.
Abstractions are easy. Tests are hard. Here's how to verify your Second-Brain OS actually works after install.
| Pillar | Test | Pass signal | Fail signal |
|---|---|---|---|
| Context | Ask "who am I?" | Reads like a teammate. | Reads like a stranger. |
| Connections | Ask "what's on today?" | Live answer in seconds. | It asks you to paste. |
| Capabilities | Type a 7-word brief. | Multi-step artifact returns. | It asks for clarification. |
| Cadence | Close laptop. Wait. | Output lands without you. | Output never arrives. |
If a pillar can't be tested, it's not in place. The framework only counts what survives a real prompt.
The install is the easy part — 20 minutes of typing. The mindset is the hard part. Get this wrong and you'll abandon the system in week two.
Three layers of AI adoption. Most people skip the first two and wonder why the third doesn't stick.
Every time you start a task — drafting a doc, replying to a thread, prepping a meeting, scanning a CSV — ask one question first: "Could AI do this for me — even 30% of it?"
30% is the magic number. You're not asking "can AI replace me on this?" You're asking "can it remove the boring part?" Half the time the answer is yes. Most of the time you discover it can do 70%. You just have to ask before you start.
Your job is a tree. The trunk is your role. The branches are the things you do — write briefs, run meetings, analyze data, manage stakeholders. The leaves are the individual atomic tasks under each branch.
Don't try to automate the trunk. You can't. Automate one leaf at a time. Pick the most boring leaf this week. Build a skill or workflow for it. Move on. Six months in, half your leaves are agent-driven and you got there one leaf per week.
AI is a mentor, not a vending machine. The vending-machine pattern is: type a question, copy the answer, close the tab. The mentor pattern is: type a question, read the answer, ask "why?", ask "what did you skip?", ask "what would you do differently if X?".
You learn ten times faster the mentor way. And you build a working model of why the system does what it does, which is what makes you good at directing it.
When you adopt this system, expect a 20% productivity drop for the first three to five days. You're learning new commands, building new habits, debugging new tools. It feels worse before it feels better.
Then it climbs. By day ten you're back to baseline. By day twenty you're 30-50% above. By day sixty you don't remember how you worked the old way.
Most people quit on day four — right at the bottom — because they were sold "instant productivity gains." It's not instant. It's a J-curve. Plan for the dip. Push through it.
The harness is opinionated. The opinions exist because the alternative — a freeform AI workspace with no rules — collapses into chaos within a month.
Six rules. They sit above any individual skill. When you're not sure what to do, return to these.
/find <topic> across Projects, Resources, Archive, and memory. The whole point of a second brain is knowing what already exists. Creating duplicates breaks that.4-Archive/, move it back. Flip its status from done to active. Don't clone a fresh project-v2. Lineage matters.CLAUDE.md. Copies drift. Single source of truth scales.0-Inbox/. Not 1-Projects/. Projects are for things that earned the scaffold. Skipping Inbox creates the unmigrated-folder graveyard the whole system exists to prevent.Boring is beautiful. Predictable beats clever. Push autonomy up only after the lower level is proven.
Every active project has exactly two meta-files. Not three. Not five.
CLAUDE.md — project context (under 60 lines): one-line summary, project-specific rules, key files, stakeholders, status frontmatter.memory.md — append-only decision log. Decisions go in. Nothing gets rewritten. This is the durable trail of why things are the way they are.Research finding from running this for months: more than two meta-files per project = bloat and abandonment within weeks. People stop opening the folder. The project goes cold. The lesson generalizes — keep the meta thin so the work can be heavy.
Five folders. That's the whole organizational schema. Inspired by Tiago Forte's PARA, simplified for an agent-native workflow.
CLAUDE.md + memory.md inside.