Super Brain
Status check 2026-11-13 — active
Verdict: The last commit is 2026-10-26 (18 days ago), PR #7 “Remove vault limits and allow material across topics”. Last release: no releases (commits only) · last commit: 2026-10-26 · licence: MIT (LICENSE file added 2026-08-25) Repo: https://github.com/superhero-tech/super-brain Changed since this note: The note (2026-11-01) describes a state the repo had already left on 2026-10-26. Commit 47d73a2 deleted
8-System/limits.md, the five failure modes the note cites as “honest limits”.
🚀 Description
superhero-tech/super-brain — “Product Builder Second Brain, AI Product Heroes Edition 3”, from the Polish team behind Superhero.tech and the AI Product Heroes course. A vault the AI reads, writes and maintains: sources compile into a linked wiki, answers get saved, projects carry enough context that an agent can join cold, and your profile lives in a file instead of a system prompt.
It is an adaptation of Karpathy Method (LLM Knowledge Bases) with the decisions already made. It is the same pattern as this vault — which makes it a handy comparison point.
Setup: git clone, cd super-brain, claude, run /second-brain-setup. Obsidian recommended but not required; also covers Claude Cowork, OpenCode and any AI IDE. Claude reads skills from .claude/skills/, Codex and ChatGPT from .agents/skills/.
đź§© Skills
| Skill | Job |
|---|---|
ingest | Lists everything a source touches, checks that reading with you before writing, then updates every touched page with two-way links (“five pages moved, or it says why not”) |
query | Answers with [source: …] citations, saves to 9-Outputs/ with a “what the wiki could not answer” section |
lint | Monthly, ten checks sorted into errors/warnings/notes, ranked report, reports only |
project-start | Interview that pushes back once per section |
project-update | Turns work into an artefact, a log entry and a fresh index row |
second-brain-setup | Profile into 8-System/about.md from an assessment (Gallup, DISC, MBTI) or interview |
youtube | Video to source note via yt-dlp, then hands to ingest |
feynman | Explains a page in plain words, sorts the thin spots into wiki’s / sources’ / own fault |
Folder numbering: 2-Inbox, 3-Projects, 7-Skills, 8-System (about.md, brain.md, limits.md), 9-Outputs.
🎨 What it adds to Karpathy’s method
Operations as skills with self-checks; a project layer (brief + rules + append-only log); a profile file; a check-in before anything is written (a wrong page today is a wrong answer for a month); frontmatter lint can sort on (status, source_count, last_updated); gaps as a required section of every saved answer; feynman; and a limits.md the agent must read.
Honest limits: error compounding is real, an authoritative-looking wiki gets checked less, and index-only navigation tops out around a hundred pages, one vault per domain. They deliberately skipped a search tool like qmd.
Reasoning for
Compared to my own setup (Brain, 400+ notes), the ideas worth borrowing: the ingest check-in (my /b:ingest only asks for cluster confirmation), the required “gaps” section in saved answers (see /b:qa), the ranked lint report, and a limits.md. My three-level index (vault-map → catalog → graph) is how I get past their 100-page ceiling. See also Second Brain Design and Progressive Disclosure.
Alternatives considered
- My own vault protocol in Brain
- Obsidian Skills — generic Obsidian skills without the project/profile layers
- Karpathy Skills — behavioural guidelines, not a vault system
đź”— Links
- Repo: https://github.com/superhero-tech/super-brain
- Companion skills: https://github.com/superhero-tech/superhero-skills
- AI Product Heroes: https://aiproductheroes.pl/
Template: tool