AI Second Brain

A repository that is three things at once: a guide to a knowledge base an agent maintains, a starter vault you copy into place, and the skills, commands and subagents that do the filing. The loop it describes is short — you clip or dump a source into raw/, an agent turns it into atomic pages under wiki/sources, wiki/concepts, wiki/entities and wiki/synthesis, links each new page into what is already there, lints what it wrote, and updates existing pages instead of piling new ones beside them. Beside the wiki sits a project layer: one folder per project with its own CLAUDE.md and an Inputs / Process / Outputs / Feedback pipeline, so the vault holds what you know and the folders hold what you are doing. It ships 18 skills, 72 slash commands, 6 subagents and a dependency-free scripts/ directory, and it runs on Claude Code over an Obsidian vault of plain markdown files. MIT licensed, 836 stars.

Who built itOne account and one repository: all 38 commits are attributed to the GitHub user undefined-ui, whose author name varies across them — undefinedKi on 19, undefined-ui on 18 and Yarchi on one — with 37 of the 38 sharing a single email address, and the contributor list holds nobody else. Nineteen commits carry a co-author trailer naming Claude Fable 5. In its first three weeks the repository reached 836 stars and 132 forks.

1. 01

   Twenty-three days, one account, half the commits naming a model

   The repository was created on 2026-09-07, its first commit initial landed that day at 15:20:52Z, and its last was 2026-09-29T15:30:38Z: 38 commits in 23 days, all inside September 2026, all from a single GitHub account. Zero releases and zero tags — the version history is the commit log, and git is documented as the maintenance mechanism (docs/09-maintenance/versioning-with-git.md) rather than consumed as tagged versions. The author name varies across them — undefinedKi on 19, undefined-ui on 18, Yarchi on one — while 37 of the 38 share one email address. 19 of the 38 commits carry a co-author trailer naming Claude Fable 5, so half the history names a model as a co-author. Around it: 836 stars, 132 forks, 7 watchers, one open issue, 2,668 KB, MIT licence, default branch main. Only three pieces of activity exist beyond the commits, and one of them comes from somebody else: issue #3, Added secondary brain, opened by cruelbites13 on 2026-09-24 and still open with no comments; issue #2, est quickdraw, opened and closed by the owner on 2026-09-23; and pull request #1, chore: normalise line endings via gitattributes, also his, with a 46-byte .gitattributes beside it. The newest commit message states the course and handbook contract and splits duplicated ground between them — the last act in the 38 was editorial de-duplication, not code.

2. 02

   The vault is the deliverable, and installing it is deleting things

   vault-template/ is 21 files, and it is the part you take away: a CLAUDE.md of 3,802 bytes at its root and a second of 654 in the example project; a wiki/ split into concepts/, entities/, sources/ and synthesis/ with an index.md (316 bytes) and a log.md (132); a projects/ layer whose example carries the pipeline Inputs / Process / Outputs / Feedback, each stage held open by an empty .gitkeep; raw/ for material you never read again; output/; and templates/ with four page types — concept, entity, source and synthesis pages, 164 to 204 bytes each. The quickstart shows how the vault is made: clone, cp -r vault-template ~/brain, copy skills/, commands/, agents/ and scripts/ into place, then delete the folder READMEs, which in the guide’s words are for reading on GitHub and not for the agent — context hygiene as an install step. Nine steps, one evening, each with a page behind it; step four has the agent interview you for the CLAUDE.md, step nine schedules maintenance. The structure is documented under docs/04-structuring/ — page types, linking rules, a frontmatter schema, naming, contradictions — but the material does not cover those pages’ contents, so the schema is unrecorded here. What the README does state is why two layers exist: the wiki holds what you know, the projects hold what you are doing, and they feed each other.

3. 03

   How it maintains itself: skills, commands and read-only subagents

   Maintaining itself is the claim in the description, and what it shows is files an agent reads rather than a service that runs. skills/ holds 18 skills, one per workflow in the guide, each a single SKILL.md of 1,689 to 3,491 bytes: ingest, chat-import, transcript, backfill, merge, lint, graph, metrics, query, report, write, publish, review, privacy, project, quiz, rename, changed-my-mind. commands/ holds 72 slash commands, 143 to 448 bytes each, and their names are the clearest statement of what the vault is expected to do to itself: /lint, /health, /stale, /orphans, /dedupe, /merge, /prune, /review, /metrics, /weekly, /schedule, /contradictions. agents/ holds 6 subagents — curator, graph-analyst, ingestor, linker, researcher, reviewer, 339 to 635 bytes each — of which the README says four are read-only by design, without naming them. The guide keeps the agent material in docs/06-agents/: roles, hooks, scheduled maintenance, safety and guardrails, subagents, skills and commands; maintenance covers linting, review cadence, git, privacy and scaling. Step nine of the quickstart is to schedule maintenance and wake up to a vault that filed itself. What the schedule runs is not in the material — neither scheduled-maintenance.md nor hooks.md was read — so the cadence, the trigger and the write scope are unrecorded.

4. 04

   Three opinions the whole thing rests on, and the case against it

   The README states three rules as the opinions everything else follows from. Nothing is ingested until it is linked, because a page that lands unconnected is invisible within a week, so linking happens in the same run or the ingest is not finished. Contradictions are recorded, never overwritten: both positions stay, with dates and sources, and the history of what you believed and why is, in its words, the one thing the vault has that a search engine does not. No vector database until you need one: a personal corpus is small and already structured, and a page covers the threshold at which RAG earns its place. A fourth rule is portability — markdown, wikilinks and SKILL.md only, with Obsidian as a viewer for the graph rather than a dependency. The lineage is plain: the LLM wiki pattern is Karpathy’s, published as a gist on 2026-04-04, and this repository is one implementation of it plus the parts the gist leaves undefined. The reading around it is the research: GraphRAG (arXiv:2404.16130), HippoRAG (arXiv:2405.14831), Lost in the Middle (arXiv:2307.03172), Matuschak’s public notes, Ahrens and Forte. What is unusual is that the strongest argument against the premise sits in the same file: if the writing is the thinking, delegating it means not doing it.

5. 05

   Three entrances, one plugin, and a 936 KB page at the repository root

   Three entrances share one repository, and its last commit was about the contract between two of them. The guide is ten sections — 65 pages in one README line and 77 in another, over 75 markdown files. The agents course is seven modules of four files each, framed on Google’s agent whitepapers and five layers: what the agent sees, who decides, who sorts work, what it can reach, how you know it works; it ships as a Claude Code plugin with its own marketplace entry, five skills, a loop-critic subagent and a two-command install. The handbooks are five tracks — knowledge graphs, Jev engineering, agent harnesses, loop engineering, eval engineering — 49 markdown files, eight or nine pages each, each ending in a build you can finish in an evening. Publishing is literal: the guide is served at undefined-ui.github.io/second-brain-os, and the built site is committed at the repository root, index.html at 936,022 bytes beside tree.html and resources.html, which is why GitHub reports the repository’s language as HTML. Two mismatches are visible from outside: the README says the repository ships four scripts while scripts/ holds six Python files, and examples/, promised as real vaults and real output, holds a README and an empty .gitkeep. CONTRIBUTING.md (978 bytes) asks of contributors what the README models itself: every factual claim needs a primary source.