GitHub - undefined-ui/second-brain-os: Guides and resources for AI engineers: an agents course, five handbooks, a hands-on second-brain project and a vetted reading list.
A free, MIT-licensed GitHub repo that teaches building AI agents through a seven-module course, five deep-dive handbooks, and a hands-on project: an agent-maintained markdown knowledge base. It also bundles a Claude Code plugin, 18 skills, 72 slash commands, and 114 curated resources.
- The course runs seven modules from a single prompt to a production agent, covering context, loops, gates, harnesses, evals and production, with a build in each module.
- The second-brain project is an agent-built knowledge base in plain markdown, set up by cloning the repo, copying a starter vault and its skills into ~/brain, and following nine setup pages.
- Its "New this month" section tracks September and October 2026 developments, including Basic Memory 1.0, Agent Skills moving to the Linux Foundation, and claude plugin eval for CI-gated plugin testing.
The repo at github.com/undefined-ui/second-brain-os is an MIT-licensed, plain-markdown learning path for AI engineers. It bundles a seven-module course on building agents (from a single prompt up to a production agent, covering context, loops, gates, harnesses, evals and production), five deeper handbooks on individual layers (knowledge graphs, gates, harnesses, loops and evals), and a hands-on project: a "second brain," a knowledge base an agent builds and maintains for you as markdown files. A Claude Code plugin ships with it, along with 18 skills, 72 slash commands and 6 subagents for the project, plus a curated list of 114 vetted resources. Nothing needs installing to read it, and the quickstart is a handful of copy commands that set up a starter vault under ~/brain, followed by nine setup pages.
The most useful part for a working engineer is the "New this month" section, which lists what changed in September and October 2026, with star counts checked on 1 November. Several items point the same direction: memory systems are converging on plain markdown with wikilinks (Basic Memory 1.0, ai-memory 3.0, agent-memory), with database-backed options like Hindsight 1.0 still in the mix. On the retrieval side, the notes flag evidence that a grep-and-BM25 harness with no vector index answered 85% of BrowseComp-Plus, and that LiteRAG claims graph retrieval over 100x faster and 99% cheaper than GraphRAG's global and DRIFT modes. Other entries cover Claude Code's new `claude plugin eval` command for running plugin cases with and without the plugin, a `/skill-doctor` command that shows per-turn context cost of loaded skills, Anthropic handing the SKILL.md standard to the Linux Foundation's Agentic AI Foundation on 22 October, and an argument from Anthropic for folding subagents back into one main loop.
Treat the claims with some care. Many of the "new" items are vendor announcements or single-paper results, and the summary itself is a curation, so the numbers (star counts, benchmark scores, percentage savings) are the repo's own reporting rather than independent verification. The practical entry point depends on your situation: beginners should read "What an agent is" and go through the course in order, engineers debugging one layer should jump to the matching handbook, and anyone who wants a project can clone the vault and work through the second-brain setup. The rest is a reading list with a stated reason for each entry, which is probably the most reusable piece for a read-it-later queue.