Build Your AI Second Brain with Claude Code

A skill for Claude Code that generates a personalized Product Requirements Document (PRD) for building your own AI second brain - a proactive, persistent AI assistant that knows your context, remembers across sessions, and keeps you informed.

From the talk: Your Second Brain: Building Personal Agents That Actually Run Your Day with Claude Code — presented by Cole Medin at the ODSC Personal AI Agents Summit 2026. This repo is the companion starter kit for that session.

What This Does

You fill out a simple requirements template describing your tools, workflow, and preferences. Then Claude Code generates a phased build plan tailored to your setup - covering memory, integrations, skills, a proactive heartbeat, chat interface, and security hardening.

The generated PRD gives you (or your coding agent) a step-by-step blueprint for building a second brain that:

• Remembers across sessions - decisions, preferences, context, all in markdown files
• Connects to your platforms - Gmail, Slack, Calendar, Asana, Linear, GitHub, and more
• Proactively monitors your email, calendar, and tasks every 30 minutes
• Chats with you through Slack (or Discord, or any messaging platform)
• Searches months of memory with hybrid keyword + semantic search
• Drafts replies in your voice using RAG on your past messages
• Tracks habits with daily nudges inspired by Atomic Habits

Quick Start

1. Install the Skill

No manual commands. From inside your own project (this can be a blank folder if you're starting your second brain from scratch), ask your coding agent (Claude Code) to fetch the skill for you - just paste this prompt:

Clone https://github.com/coleam00/second-brain-summit into a folder right next to our current project, then bring its create-second-brain-prd skill into .claude/skills/ in our current working directory so it's available here.

The agent clones the repo alongside your project and drops the skill into .claude/skills/create-second-brain-prd/. That's the whole install.

2. Fill Out the Requirements Template

Copy the template to your workspace and fill it out:

cp .claude/skills/create-second-brain-prd/my-second-brain-requirements.md \
   ./my-second-brain-requirements.md

The template has 8 sections:

1. About You - Name, role, timezone
2. Your Platforms - Which tools you use (Gmail, Slack, Linear, etc.)
3. Top Tasks - 3-5 things you want AI to handle proactively
4. Proactivity Level - Observer, Advisor, Assistant, or Partner
5. Security Boundaries - What the agent should never do without permission
6. Memory Categories - What types of knowledge matter to you
7. Infrastructure - OS, local vs. VPS deployment
8. Integration Priority - Which 3 integrations to build first

See example-my-second-brain-requirements.md in the skill directory for a completed example.

3. Generate Your PRD

Open Claude Code in your project and run:

/create-second-brain-prd ./my-second-brain-requirements.md

Claude will:

1. Read your requirements
2. Load the architecture reference blueprint
3. Research every tool and API in your stack via web search
4. Generate a personalized 9-phase PRD at .agent/plans/second-brain-prd.md

4. Build It

Follow the phases in your PRD. Each phase includes:

• What to build and why
• Key files to create (with paths)
• Dependencies on previous phases
• Complexity estimate
• Personalization notes based on your answers

The recommended build order:

Phase	What	Complexity
1	Memory Layer (BOOTSTRAP.md, SOUL.md, USER.md, MEMORY.md, daily logs)	Low
2	Hooks (SessionStart, PreCompact, SessionEnd)	Medium
3	Memory Search (hybrid keyword + semantic)	Medium
4	Integrations (your top 3 platforms)	Medium each
5	Skills (vault structure + custom skills)	Low-Medium
6	Proactive Systems (heartbeat + daily reflection)	High
7	Chat Interface (Slack/Discord bot)	High
8	Security Hardening (sanitization, guardrails)	Medium
9	Deployment (local scheduler or VPS)	Medium

Starter Templates

This repo ships a ready-to-use templates/memory/ folder so you don't start from a blank page. Copy it into your vault and customize:

• BOOTSTRAP.md - a first-run onboarding wizard. On your very first session it interviews you one question at a time (name, role, timezone, communication style, integrations, proactivity), fills in USER.md, SOUL.md, and HEARTBEAT.md for you, then deletes itself.
• SOUL.md - how your agent behaves: identity, values, communication style, boundaries.
• USER.md - about you and your company: profile, team, accounts, preferences.
• MEMORY.md, HEARTBEAT.md, HABITS.md, daily/ - the rest of the memory layer.

cp -r templates/memory your-vault/Memory

Architecture Overview

The second brain is built on Claude Code and the Claude Agent SDK. No massive framework - just markdown files, Python scripts, and an Obsidian vault.

🗺️ Complete Workshop Reference (the full walkthrough)

The diagram above is the high-level view. The one below is the exhaustive reference from the full workshop - it expands every subsystem into its own flow: the 3-layer security pipeline, the memory-search pipeline (chunk → embed → index → hybrid search), the heartbeat loop, vault sync, and the local vs. VPS infrastructure split.

Memory Layer (center of everything)
    BOOTSTRAP.md - First-run onboarding wizard (writes SOUL/USER, then self-deletes)
    SOUL.md - Agent personality, values, boundaries
    USER.md - Your profile, accounts, preferences
    MEMORY.md - Key decisions, lessons, active projects
    daily/YYYY-MM-DD.md - Timestamped session logs

Hooks (context persistence)
    SessionStart - Loads memory into every conversation
    PreCompact - Saves context before auto-compaction
    SessionEnd - Captures decisions on exit

Integrations (platform connections)
    Python CLI wrapper pattern - LLM never sees API keys
    query.py gmail list / query.py asana overdue / etc.

Skills (extensible capabilities)
    Progressive disclosure - metadata always loaded, full instructions on demand

Heartbeat (proactive monitoring)
    Python gathers data -> Claude reasons -> notifications sent
    ~$0.05/run vs $0.38 with MCP tool calls

Memory Search (hybrid RAG)
    FastEmbed (local ONNX) + SQLite/Postgres
    70% vector + 30% keyword = best of both worlds

Proactivity Levels

Your choice in the requirements template shapes the entire system:

Level	What It Does
Observer	Notifications only. Never takes action.
Advisor	Drafts emails/messages for your review. Tracks habits with suggestions.
Assistant	Auto-organizes files, auto-logs decisions. Asks for anything external.
Partner	Sends low-risk messages, completes routine tasks. Asks only for irreversible actions.