A second brain for product managers. Plain markdown files in a folder on your laptop. Claude reads them before answering, writes to them after, sweeps them every Friday. No vector DB. No cloud. No agent memory tricks.
You manage one product. Your context is scattered across Notion, Linear, Slack, your dashboards, and your head. You ship a feature. Six weeks later, nobody remembers why you killed the other option. PM Brain fixes that.
Research preview. The architecture has months of dogfooding behind it on a sister content brain; the product as installed by real PMs in real organizations is days old. The eval suite (404 of 406 checks pass) is the floor. Real-world install feedback is how the next version gets better. Open issues or join the conversation per the launch post.
Install
Two stages: install the skill once, then bootstrap a brain in any folder you want.
Stage 1: install the skill (one time)
The skill is global. It lands in and is available across every Claude Code session, in any working directory.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Either path pulls only the skill folder (.claude/skills/pm-brain/) into your Claude Code skills directory. The rest of this repo (example-brain/, tests/, docs/) stays on GitHub for you to browse, not on your laptop.
Stage 2: bootstrap a brain (per-product)
In any folder where you want the brain to live:
cd ~/projects/my-product-brain
claude
# then in the Claude Code prompt:
/pm-brain
The skill detects what's already in the directory. An empty folder gets a fresh start (greenfield). A folder with existing PM artifacts (Notion exports, a Jira CSV, meeting notes) gets read and absorbed (migration). Either way, a short 5-batch interview captures company, role, and current priorities. The scaffold drops in, the brain commits locally. Never pushes.
Migration is for your current state. Don't backfill old artifacts retroactively. Migration mode reads your active strategy, in-flight hypotheses, recent decisions, current stakeholder list. That's the goal. Don't try to manually /ingest 200 old interview transcripts and six months of Slack threads. If a stale artifact matters, it'll come up through current work and you'll ingest it then. Forcing them in now wastes a weekend and clogs the durable layer.
Stuck? Universal fallback (any OS with git): git clone https://github.com/phuryn/pm-brain.git && cp -R pm-brain/.claude/skills/pm-brain ~/.claude/skills/. On Windows replace the cp -R with Copy-Item pm-brain\.claude\skills\pm-brain $env:USERPROFILE\.claude\skills\ -Recurse.
What it does (one loop)
The brain has one loop. Every task runs through it.
Ingest. Feed it an artifact (transcript, doc, screenshot, line in chat, app-connector pull from Notion / Jira / Slack).
Source + synthesize. The original copies to source/ (immutable). The synthesis lands in ingestion/ with observations tagged by speaker and date.
Propagate. The durable layer (knowledge/, hypotheses/, decisions/, stakeholders/) gets updated wherever the new signal applies. One artifact often touches four to six files.
Tag. Every load-bearing claim wears a provenance marker: a documented interview, a verbal stakeholder comment, your hunch, or general industry knowledge. The tags carry an implicit hierarchy: documented outweighs verbal, verbal outweighs intuition. The weighting is in plain text, so you can see when the brain is leaning on it and override it when your judgment says otherwise.
Sweep. Friday's /review reads the whole folder, flags what's drifting, drafts what needs your call.
Full list of provenance tags in the glossary. Worked end-to-end example: how it works.
The six commands
| Command | What it does |
|---|---|
| /ingest <thing> | Feed an artifact into the brain: a file, a paste, a quick note in chat. The skill figures out the shape (interview, meeting, market signal, ad-hoc) and routes it. |
| /prep <stakeholder> | One-page brief before a meeting: their open asks, last unresolved concern, suggested questions. |
| /review | Weekly sweep. Six checks across the brain. Fixes small things directly, drafts the bigger ones for your call. |
| /ideate <problem> | Synthesis, not brainstorm. Loads strategy, insights, and hypotheses. Surfaces 3–7 directions, each tagged with the evidence behind it. |
| /risk <feature> | Five-area risk scan. Drafts hypothesis stubs for any area with no coverage. |
| /plan <objective> | Six-block draft plan: what we know, assumption vs evidence, who to interview, hypotheses to open, experiments to run, decision points. |
What's in the repo
.claude/skills/pm-brain/ # The canonical skill. Install this.
example-brain/ # Pre-scaffolded instance (browseable demo)
tests/ # Eval suite. Synthetic scenarios + harness.
docs/ # Architecture, how it works, testing, prior art
The brain itself (once you install and run /pm-brain) lives in your working directory:
knowledge/: your stable picture of strategy, product, users, market, and org
hypotheses/: things you're tracking the evidence for (e.g., "invite-link friction blocks team activation under 50 seats")
decisions/: calls you've made, with the evidence trail and what would reopen them
stakeholders/: one file per person, with their asks and concerns
ingestion/: synthesis of every interview, meeting, doc, or message the brain has read
source/: the untouched originals, so the audit trail stays intact
What it isn't
Not a notes app. It's an opinionated structure with maintenance built in.
Not a chatbot with memory. The memory lives in your repo, not in Claude.
Not a vector database. Every file is plain markdown, grep-able by you.
Not an agent memory system. Nothing is embedded, nothing is retrieved by similarity. The brain stores only what you and the agent deliberately wrote down.
Not autonomous product management. Judgment stays with you. The brain makes the boring cross-referencing easier.
Tests
17 synthetic PM scenarios. 404 of 406 individual checks pass across the snapshots (≈99.5%). The split: every structural check passes (329 / 329, 100%). Files exist, links resolve, evidence rows tagged, decision schemas valid. And 75 of 77 LLM-judge content checks pass (≈97%), with two judges missing on the two longest scenarios. Full breakdown in the scoreboard.
Each scenario is a multi-turn PM situation (a churn investigation, a stakeholder cadence flag, a contradiction arriving 60 days after a decision) with cached input artifacts and ground-truth assertions. The harness spins up a fresh brain in a temp dir, replays the inputs through claude -p, runs structural assertions after every turn, and runs LLM-judge rubrics on substance at scenario end.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
