phuryn/pm-brain

PM Brain is a second-brain system for product managers built on plain markdown files in a folder on a laptop. Claude Code reads the files before answering, writes back into them after, and runs a Friday sweep across the whole folder. There is no vector database, no cloud sync, and no embedded agent memory. The author frames it as a research preview: the architecture has months of dogfooding behind it on a content-focused sister project, but real-team PM installs are days old. Install has two stages. First, a one-time install of the skill itself into ~/.claude/skills/pm-brain/, which the README shows via a curl-and-tar one-liner on macOS, Linux, WSL, and Git Bash, plus a PowerShell equivalent for Windows. Second, a per-product bootstrap: cd into the folder you want the brain to live in, run claude, and call /pm-brain. The skill either starts greenfield in an empty folder or absorbs existing PM artifacts like Notion exports and Jira CSVs in migration mode. A short five-batch interview captures the company, role, and current priorities. The brain has one loop. You ingest an artifact, the original goes into source/ untouched and a synthesis lands in ingestion/, the durable layer (knowledge/, hypotheses/, decisions/, stakeholders/) is updated wherever the signal applies, and every claim is tagged with provenance: documented interview, verbal comment, your hunch, or general industry knowledge. Friday's /review reads the whole folder, fixes small drift, and drafts the bigger calls. Six commands drive day-to-day use: /ingest feeds in an artifact, /prep writes a one-page brief before a stakeholder meeting, /review is the weekly sweep, /ideate surfaces three to seven directions with evidence tags, /risk runs a five-area scan, and /plan drafts a six-block plan covering what is known, assumptions, who to interview, hypotheses, experiments, and decision points. The README is explicit about what PM Brain is not: not a notes app, not a chatbot with memory, not a vector database, not an agent memory system, and not autonomous product management. Judgment stays with the human, the brain only stores what the user and agent deliberately wrote down. The eval suite reports 404 of 406 checks passing across 17 synthetic PM scenarios, roughly 99.5 percent, with all 329 structural checks green and 75 of 77 LLM-judge content checks passing.