second-brain-starter: A PRD Generator for a Persistent Claude Code Assistant

The output is a plan, not an assistant

second-brain-starter is a skill for Claude Code, and what it produces is a Product Requirements Document rather than a running assistant. The workflow starts with a requirements template describing your tools, workflow and preferences, and Claude Code then generates a phased build plan tailored to that setup, covering memory, integrations, skills, a proactive heartbeat, a chat interface and security hardening. The document lands at `.agent/plans/second-brain-prd.md` as a nine phase PRD. The generated PRD is described as a blueprint you follow yourself or hand to a coding agent, and each phase includes what to build and why, the key files to create with their paths, dependencies on earlier phases, a complexity estimate and personalization notes drawn from your answers. In other words, you are given a spec and a sequence, and the building is yours to do.

Eight questions that determine the whole system

The requirements template is short, and the section list is the clearest statement of the design space. Section one is About You, covering name, role and timezone. Section two is Your Platforms, meaning which tools you use, with Gmail, Slack and Linear given as examples. Section three is Top Tasks, where you name three to five things you want the assistant to handle proactively. Section four sets the Proactivity Level. Section five is Security Boundaries, describing what the agent should never do without permission, which is the section worth filling in most carefully. Section six is Memory Categories, what kinds of knowledge matter to you. Section seven is Infrastructure, the operating system and whether you deploy locally or on a VPS. Section eight is Integration Priority, which three integrations to build first. A completed example sits in the skill directory as `example-my-second-brain-requirements.md`.

Installing the skill and generating the PRD

Installation is a copy of a directory. You clone the repository and move the skill folder into your own project.

# Clone this repo
git clone https://github.com/coleam00/second-brain-starter.git

# Copy the skill into your project
cp -r second-brain-starter/.claude/skills/create-second-brain-prd \
      your-project/.claude/skills/create-second-brain-prd

Alternatively you can copy the whole `.claude/` folder from the repository into an existing project. Then you copy the requirements template into your workspace and fill it in.

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

With Claude Code open in your project, you invoke the skill.

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

Claude then reads your requirements, loads the architecture reference blueprint, researches every tool and API in your stack through web search, and writes the personalized nine phase PRD. The research step is why the plan can name actual APIs rather than generic placeholders.

The nine phases, in order and by cost

The recommended build order is given with a complexity rating for each phase, which is a useful calibration for anyone deciding what to attempt first. Phase one is the memory layer, covering BOOTSTRAP.md, SOUL.md, USER.md, MEMORY.md and daily logs, rated low. Phase two is hooks for SessionStart, PreCompact and SessionEnd, rated medium. Phase three is memory search using hybrid keyword and semantic retrieval, medium. Phase four is integrations for your top three platforms, medium each. Phase five is skills, a vault structure plus custom skills, low to medium. Phase six is the proactive systems, a heartbeat and daily reflection, rated high. Phase seven is a chat interface, a Slack or Discord bot, also high. Phase eight is security hardening with sanitization and guardrails, medium. Phase nine is deployment, a local scheduler or a VPS, medium. The order is not arbitrary: the memory layer and hooks come first because everything later depends on having somewhere to put what the agent learns.

The memory vault is markdown files

The architecture overview is unusually concrete about the storage layer. The second brain is built on Claude Code and the Claude Agent SDK with no large framework, just markdown files, Python scripts and an Obsidian vault. The memory layer sits at the centre. BOOTSTRAP.md is a first-run onboarding wizard that, on your very first session, interviews you one question at a time covering name, role, timezone, communication style, integrations and proactivity, fills in USER.md, SOUL.md and HEARTBEAT.md for you, and then deletes itself. SOUL.md holds the agent's identity, values, communication style and boundaries. USER.md holds your profile, team, accounts and preferences. MEMORY.md, HEARTBEAT.md, HABITS.md and a daily folder make up the rest, with daily logs timestamped per session. The starter templates folder ships these so you are not writing them from scratch.

cp -r templates/memory your-vault/Memory

Hooks, heartbeat, search and habit tracking

Three mechanisms make the assistant persistent rather than a chat window. Hooks handle context persistence: SessionStart loads memory into every conversation, PreCompact saves context before automatic compaction, and SessionEnd captures decisions on exit. The heartbeat handles proactive monitoring, and the architecture note describes the split of labour as Python gathering data, Claude reasoning, and notifications sent afterwards. That design comes with a stated cost of about $0.05 per run, against $0.38 using MCP tool calls, which is a large enough difference to change the shape of what you monitor. Integrations follow a Python CLI wrapper pattern where the LLM never sees API keys, so credentials stay in the wrapper rather than in a prompt. The skill layer uses progressive disclosure, with metadata always loaded and full instructions fetched on demand. Retrieval is described as hybrid RAG, with FastEmbed running locally on ONNX and storage in SQLite or Postgres, weighted at 70 percent vector and 30 percent keyword. The stated rationale is having both, since pure vector search loses exact identifiers and pure keyword search misses paraphrase. Search is meant to cover months of memory rather than the current conversation. On top of retrieval sits a drafting feature: the assistant writes replies in your voice using RAG over your past messages, which is the part that depends most on having a long enough history before it produces anything useful. Habit tracking uses daily nudges inspired by Atomic Habits and lives in HABITS.md. The proactivity levels decide how much of this happens without you, which is the parameter that actually changes the risk profile of the whole system.

Observer, Advisor, Assistant, Partner

Four levels are defined, and they are the clearest way to describe what the system is authorised to do. Observer produces notifications only and never takes action. Advisor drafts emails and messages for your review and tracks habits with suggestions. Assistant goes further, auto-organizing files and auto-logging decisions, while asking for anything that reaches outside. Partner sends low-risk messages and completes routine tasks, asking only for irreversible actions. Your choice in the requirements template shapes the entire system, which means the template is where you set the boundaries rather than the code. Security hardening is a later phase in the plan, covering sanitization and guardrails, and the requirements template asks separately what the agent should never do without permission. Two of the four levels, Assistant and Partner, act without asking, so the risk of a misconfigured agent rises sharply between them.

Licence, releases and what the repository does not settle

Two facts are missing from the repository and both matter. There are no GitHub releases, so there is no version to pin, and the repository is not archived with the last push on 2026-06-09, so the skill tracks whatever the files currently contain. The licence is not declared, which is unusual for a repository people are asked to copy into their own projects and modify. Anyone redistributing a customised version, or shipping one inside a product, has to sort out the terms first. The repository itself is small: a `.claude/` directory holding the skill, a `templates/` folder with the memory layer, and two images that explain the pitch. A community workshop and links to the Claude Agent SDK documentation and Obsidian are offered as further reading, so the project is a starting point for a workshop exercise rather than a maintained product.

Editorial conclusion

second-brain-starter fits an individual who wants a personal assistant designed around their own tools and boundaries, and who is willing to spend a session answering questions before any code exists. It is the wrong choice for a team that needs a shared, supported product, since the repository publishes no releases and states no licence, so the terms for redistribution are unclear. Before you start, copy the requirements template to your workspace and fill in all eight sections, because the generated phases are only as specific as the answers you give it.

Frequently asked questions

Official sources

1. coleam00/second-brain-starter on GitHub
2. Issues
3. README