Read this in other languages: Русский
A second brain and persistent memory for Claude Code and Codex: 234 skills shipped as a plugin marketplace, a memory server you can attach with one manifest, a working CRM engine and vault templates. It is the system Palo Alto AI Research Lab runs every day, stripped of personal data and cut down to a portable core: the method, not the data.
Pick your door:
- New to this? Start with the light onboarding guide. Download a folder, open it in Claude Code or Codex, and try one useful task. No terminal, GitHub account, shared vault, sync service, or background updater is needed; your work stays in a separate local folder.
- Already have an Obsidian vault? Take the habits that keep it from rotting without the rest:
the concept template with a
maintained_by: humanflag so agents never rewrite your notes, three read-only subagents that review and propose, and the contract for unattended runs. - Already running agents across machines? Read HANDOVER.md: a snapshot of the whole system and a map of our repos, written for you and your model.
This is an optional next stage. First-time users should finish the light route before installing the full kit below.
A single skill is a trick. What makes an agent useful over months is the set of habits around it: a gate that refuses to say "done" without evidence, a retrospective that files what you learned where the next session will actually find it, a registry of decisions you already rejected so nobody re-pitches them, a handoff that survives a crash. Those only work together, which is why they ship as one kit rather than as 234 separate downloads.
Install it whole and you get a working second brain on day one:
npx skills add tonydzi/second-brain-starter-kitor, inside Claude Code:
/plugin marketplace add tonydzi/second-brain-starter-kit
/plugin install second-brain-skills@second-brain
Want the agent to remember across sessions? Memory is a separate, opt-in plugin in the same marketplace, because its embeddings stack downloads about 2 GB on first start:
/plugin install second-brain-memory@second-brain
It registers the sqlite-graph-memory MCP
server (sgm-mcp) and asks for one setting, the index folder (BRAIN_INDEX_DIR) you built with
uvx --from 'sqlite-graph-memory[embeddings]' sgm-index <your notes folder>.
It is MIT, it costs nothing, and nothing here phones home. Take what you need, delete the rest, and open an issue if a skill breaks on your machine. Paths are documented in docs/PATHS.md.
Setting up a brand-new computer, not just adding skills? There is a full runbook a fresh Claude Code session can execute end to end — the human says one sentence, the agent does the rest:
- macOS → ONBOARDING-MAC.md
- Windows → ONBOARDING-WINDOWS.md
- joining an existing fleet as a trusted family node (sync, shared memory, message bus) → ONBOARDING-FAMILY.md — method only; fleet coordinates travel over the encrypted sync channel after the owner approves the device, never through this repo.
Every one of them ran in production before it was published. One command each.
| skill | what it gives you in 60 seconds |
|---|---|
/secondop |
a second opinion from a PANEL of external LLMs (Codex + Grok + Gemini) on any plan or diff — a model never reviews itself |
/tt |
the quality gate after every build: run it live, break it on purpose, prove it with a counter — only then say "done" |
/retro |
end-of-session retrospective that routes every durable artifact to its home and hands you a compact handoff |
/handoff |
a self-contained "semicolon" so another session, person or machine continues exactly where you stopped |
/skill-gap |
audits your sessions for repeated manual patterns and tells you which skill to install or build next |
/precedent |
before you decide anything structural: did we already decide it, and did we already reject it? |
/declined |
the registry of rejected ideas, so the agent stops re-pitching what you already said no to |
/issue-match |
measures a target repo's queue before you spend a PR: how many of the last closed PRs were actually merged |
/research-swarm |
turns any hypothesis into an argument map from 5 independent lenses — never a verdict |
/last30days |
deterministic "what changed in the last 30 days on topic X" before a strategic call |
/ask |
semantic search over your curated vault — the smallest relevant slice, not a corpus dump |
/search |
full-text search across every conversation you ever had, locally, 0 tokens |
/chat-search |
find which past Claude Code session discussed a thing — and resume it |
/resume-last |
continue the previous chat after a crash, even on another machine where native resume can't reach |
/1 |
hard-crash recovery: where were we + is everything alive + restore the history, in one command |
/arch |
the deterministic map of everything your system is made of, and what fell off it |
/brain |
one-glance health of a memory/RAG stack so failures stop being silent |
/sync-check |
green/red report on file sync across a machine fleet, including the silent sync-conflict count |
/quarantine |
holds unsigned or unverified incoming deliverables instead of applying them — fail-closed |
/obsidian-ingest |
universal import pipeline: any source becomes well-linked atomic notes with provenance |
/defuddle |
clean web→markdown without ads or nav — 40-60% fewer tokens than a raw fetch |
/dedup |
find and merge near-duplicate notes with a supersede-not-delete policy |
/taste-check |
a content quality gate: is this fit to show, before anyone sees it |
/intake |
one pass routes a new rule into every home it needs, so parallel sessions see it on their own |
/mcp |
health-check your MCP servers and scout new ones before any integration work |
If one of these saved you an hour, a star is the thanks we would ask for: it is how the next person searching for agent memory finds the kit.
Standalone tools (separate repos, same lab):
| tool | what it does |
|---|---|
| claude-memory-tidy | keeps an agent memory folder clean and deduplicated |
| sqlite-graph-memory | graph memory for agents on plain SQLite — no server, no vector DB |
| verbatim-citation-gate | a deterministic gate that catches invented quotes before they ship |
| agent-leash | an 8-domain leash for delegated agent authority |
| claude-consensus | machines negotiate propose → counter → accept → commit instead of drifting |
| llm-spend-audit | where your LLM subscription budget actually goes |
| agent-runtime-integrity-bench | replay, idempotency and consensus-integrity scenarios from real incidents |
Full map of all 234 → skills/INDEX.md
| Folder / file | What it is |
|---|---|
SEED.en.md / SEED.md |
The opening message for your FIRST Claude Code session (English / Russian) — everything starts here |
BOOTSTRAP-CLAUDE.md |
Instructions for Claude itself: how to install and adapt the kit |
CLAUDE-EXTERNAL.md |
Our working principles (assistant behavior) — the base for your own CLAUDE.md |
skills/ |
All 234 skill commands the system runs on every day, map: skills/INDEX.md. Personal data in examples is replaced with fictional stand-ins (how exactly) |
templates/ |
Second-brain note templates (concepts, decisions, weekly/monthly reviews) |
crm-template/ |
CRM: markdown "one card = one file" + an engine with warmth scoring and safe outbound, with demo data |
docs/ |
How to use CLAUDE.md, onboarding, what is shared |
agents/ |
Three read-only subagents (reviewer, curator, graph-analyst): they report and propose, never write |
docs/SCHEDULED-RUNS.md |
The contract every unattended run follows: a log line, one commit, one line to you, and how to catch a run that silently stopped |
RED-PROOF.md / llms.txt |
A drop-in rule for any CLAUDE.md or AGENTS.md: a check counts only after you saw it fail / a one-page summary of this repo for LLM readers |
HANDOVER.md |
Already running your own agent fleet? A snapshot of the whole system and a map of our repos — the entry point for you and your model |
📖 Repository wiki: what is actually in here · architecture (4 layers) · your first week · adapting it to your machine
Don't need the whole kit? Install just the skills, straight into your Claude Code:
/plugin marketplace add tonydzi/second-brain-starter-kit
/plugin install second-brain-skills@second-brain
Or via skills.sh (works for Claude Code, Codex, Cursor, Gemini CLI and any other agent that speaks the Agent Skills format):
npx skills add tonydzi/second-brain-starter-kit
- A Claude Pro or Max subscription → install Claude Code:
npm install -g @anthropic-ai/claude-code(no npm? Mac:brew install node, Windows: nodejs.org). - Run
claudeand log in with your subscription. - Install Obsidian — to look at your second brain with your own eyes.
- Clone this repository:
git clone https://github.com/tonydzi/second-brain-starter-kit.git(or Code → Download ZIP). - Open
SEED.en.md, put your name in, and paste the text as the first message to Claude Code. Claude takes it from there, followingBOOTSTRAP-CLAUDE.md. (Russian version:SEED.md.)
Using Gemini CLI instead of Claude? → docs/INSTALL-GEMINI.md
- A second brain = stop losing what's useful. Every decision, agreement and idea → a note in the vault, linked into the graph.
- AK-47 rule: the simplest solution the owner can repair themselves.
- Receipts: the assistant proves it saved something ("recorded X → note Y") instead of saying "understood".
- A routine repeated twice becomes a skill — a
/namecommand. - Privacy: everything lives locally on your machine; money, deletion and anything leaving the house happen only on your explicit OK.
Now: v0.9.0.
234 skills, 247 engine scripts under engines/, note templates, a CRM engine with warmth
scoring and safe outbound, SEED.md + BOOTSTRAP-CLAUDE.md for the first session, and
HANDOVER.md for people who already run their own agent fleet.
Next, in the order we would take it ourselves:
- Try the 30-minute onboarding route through the shipped rules, skills, engines, templates and CRM demo; report any step that does not work on your machine.
- A "clone it and it works" check on someone else's machine. Right now the only proof is that we use this ourselves every day. That is usage, not a test: no third-party machine has ever been seen.
- Localized READMEs for non-English venues (Chinese, Japanese, Korean, Spanish) — the skill bodies stay English, the docs get localized.
Every notable change ships as a release, so the releases feed is an honest record of what is actually in the kit — unlike a commit history.
This project is built by a human + AI team, and the commit history shows it: Claude writes most of the code, Codex and Grok review it, Gemini brings research. A model is credited in a commit only if its output actually changed that commit's content — no decorative signatures. Lab-wide rule: AI-CONTRIBUTORS.md.
The method is open — take it, adapt it, share it. There are no secrets and no personal data in the kit (it goes through an automated secret scanner). Questions and ideas → Issues.
Questions, war stories, or you want to run this yourself:
- 💬 WhatsApp: +1 341 222 9178
- 🐦 X: [аккаунт]
- 📣 Telegram: @ClawEng (EN) · @ClawRus (RU)
- 🌐 palo-alto.ai · Palo Alto AI Research Lab
We spend a lot of time answering other people's issues. It was fair to say out loud what we have not built ourselves:
- Clone this on a machine that is not ours and write down everything that breaks
- Localize the README (Chinese, Japanese, Korean or Spanish)
Issues labelled accepted are scoped, free to take, and nobody is on them.
Comment "claiming this" — no permission needed — and it is yours for 7 days.
New here? Start with good first issue.
You keep the copyright to your code. No CLA, no assignment, ever — your contribution goes in under this repo's existing license, the same terms as ours. We answer every issue and PR within 48 hours, including "no, and here is why"; our silence is our bug, so ping the thread.
Full deal: CONTRIBUTING.md
Live on a platform we do not use daily (Codex, Cursor, Gemini CLI, Windows and others)? Own it: PLATFORMS.md lists who owns what, and the pinned issue #10 "Own a platform" is where you claim one. How we review and credit contributions in this repo: CONTRIBUTING.md.
This repository is one piece lifted out of a live operation: one engineer running a fleet of machines, built with Claude as implementation collaborator; the machines reach consensus with each other and wake the human only for money or the irreversible. It was extracted after it survived production, not written as a demo — and it runs on its own: nothing here phones home to the rest.
See how the whole thing fits together → SYSTEM.md
Its closest neighbours in the memory layer: voice2brain · sqlite-graph-memory
Built and battle-tested at Palo Alto AI Research Lab — a fleet of Claude Code machines running 24/7 as a second brain and synthetic cofounder. Every skill here survived real production use before publication.
- 📦 All 234 skills: https://github.com/tonydzi/second-brain-starter-kit
- 👤 Author: Anton Dziatkovskii — Telegram @tonydzi · WhatsApp +1 341 222 9178 · X [аккаунт]
- 🧪 Engineers: want to test-drive this setup? Message me — I hand out free starter seeds to engineers who test and report back. Custom skill requests welcome.
One click and an agent reads the repo, pulls out the patterns and helps you apply them to your own work.
Copy the prompt (works in any agent: Gemini, Grok, a local model, your own CLI)
Read this repo: https://github.com/tonydzi/second-brain-starter-kit (“second-brain-starter-kit” - Claude Code as a second brain: 234 battle-tested skills, a working CRM engine, vault templates and the handover map. Method, not data. MIT). Work out what problem it actually solves, pull out the reusable patterns and help me apply them to my own setup. Start by asking what I am working on.
— TonyDzi, Palo Alto AI Research Lab · second brain, agent coordination, persistent memory: github.com/tonydzi