KP-llm-BrainKit

A Claude + Obsidian second-brain starter kit. It gives you a persistent, compounding knowledge vault and a set of disciplined skills that make Claude plan, build, debug, orchestrate, and file knowledge the same way every session, instead of re-deriving everything from scratch each time.

It ships empty. You bring your own projects and notes. Nothing in here belongs to anyone else.

Based on Andrej Karpathy's LLM-wiki pattern (gist 442a6bf555914893e9891c11519de94f).

What you get

• A blank Obsidian vault with a battle-tested schema: an index, an append-only log (with size-based rotation to log-archive/), and folders for sources, entities, concepts, syntheses, lessons, and projects. The CLAUDE.md inside it is the operating manual Claude reads first every session: session boot rules, a skill routing table, linking hygiene, recency/confidence conventions, and the four discipline rules that keep the vault lean.

• Twelve skills that auto-trigger on natural phrases:

  • KP-Grill — grills your idea with one sharp question at a time, files a PRD.
  • KP-Setup — scaffolds a new project (folder + matching vault pages) and audits existing ones.
  • KP-Migrate — retrofits an existing code project into the system (credentials by reference only, never copied).
  • KP-BugFix — a disciplined six-phase debugging loop that files a post-mortem.
  • KP-WikiHealth — full-vault health scan with a mechanical Python scanner: dead links, orphans, stale pages, duplicates, index drift.
  • KP-God — the orchestrator: turns a chat into a thin conductor running several parallel workstreams, each as a real Claude Code session in its own WezTerm tab with an isolated git worktree, steered across chunks and fix rounds, with a zero-token watcher and a per-project board file as its memory. Windows-native fleet; see prerequisites below.
  • code-cowork — pair-programming build mode: plan, architect against existing ADRs, trace the runtime read path for config changes, build with TDD, then two end gates (a deterministic fallow audit and an adversarial semantic self-review).
  • check-pr — review a GitHub/GitLab PR against the plan and the codebase.
  • improve-codebase-architecture — find consolidation and refactoring opportunities, report-only.
  • wrap-up — end-of-session sweep: quality pipeline per touched project, vault lint, index/log/status updates.
  • handoff — compact a conversation into a doc the next session picks up.
  • signoff — one-command session close: handoff, then wrap-up, then a copy-paste prompt to continue in a fresh chat.

• Three agents for KP-God's fleet, installed to ~/.claude/agents/:

  • codis — senior implementer: reads the codebase deeply, reuses before inventing, plans-and-builds a lane in one coherent context.
  • revis — adversarial reviewer: cross-checks the diff against the plan's acceptance criteria, hunts reinvention, dead code, and silent no-ops. Read-only.
  • planner — high-stakes planning specialist for irreversible lanes that need your sign-off before any code.

• The matching slash-commands (/KP-Setup, /KP-Grill, /KP-Healthcheck, /KP-Migrate, /KP-BugFix, /KP-WikiHealth, /KP-God, /code-cowork). wrap-up, handoff, and signoff trigger on natural phrases ("wrap up", "sign off") rather than a slash-command.

• An installer that copies everything into place, writes the one config file (~/.claude/brainkit.json: your name + vault path), pulls three official Anthropic plugins (superpowers, frontend-design, playground), and checks the KP-God prerequisites on Windows.

Prerequisites

Core kit (all platforms):

• Claude Code
• Obsidian (free)
• git

Optional but recommended for code projects (the skills use them when present, skip them when absent):

• fallow — deterministic JS/TS audit gate: npm install -g fallow
• graphify — code knowledge graph: uv tool install graphifyy (or pip install graphifyy)

KP-God's lane fleet (Windows only) additionally needs:

• WezTerm — winget install wez.wezterm
• Claude Code CLI on native Windows (claude --version works in PowerShell)
• git 2.5+ (worktrees) and Node.js

On macOS/Linux everything except the KP-God fleet scripts works; the orchestration skill's concepts still read fine, but lanes-in-WezTerm-tabs is Windows-native.

Quickstart

1. Install Obsidian (you do not open a vault yet — the installer creates one).
2. Get the kit. Clone it, or just give Claude Code the link to this repo:

   git clone https://github.com/<this-repo-owner>/KP-llm-BrainKit.git
   

3. Tell Claude install this from inside the repo. The install skill asks your name and where to put your Brain folder, copies the skills, commands, and agents, scaffolds the vault, writes the config, and checks the KP-God prerequisites.

Prefer to do it by hand? See INSTALL.md — it includes the full fresh-PC setup for KP-God.

After install: open the Brain folder in Obsidian, read its CLAUDE.md once, then say KP setup to scaffold your first project — or drop a file in raw/ and say ingest.

How the pieces fit

• One feature, built start to finish → code-cowork inline. One agent, one context, the quality gates at the end.
• Something broken → KP-BugFix. Reproduce first, fix last, post-mortem filed.
• Fuzzy idea → KP-Grill until it is a PRD, then build.
• Several genuinely parallel streams → KP-God. The conductor stays thin, lanes run as visible Claude sessions you can watch and type into, revis reviews every code lane, and the board file survives any chat dying.
• End of the day → wrap-up (or signoff if the next chat should continue the work).

What this is NOT

• It is not a copy of anyone's vault. It is a clean template; your content starts empty.
• It does not include any business data, credentials, or personal notes.
• It does not wire up any paid review tools. The skills run on Claude Code alone, plus the two optional free CLIs above.

License

No license yet. Until one is added, treat this as "all rights reserved" for redistribution; you are welcome to use it for your own setup.