相棒

August 13, 2026 · View on GitHub

Who writes this and why → START-HERE.md · one page of proof: tonydzi.github.io

THE JOURNEY with Claude Code

How one non-technical founder and an AI built a second brain, day by day.

This is a build-in-public book. Anton Dziatkovskii started working with Claude Code on 2026-05-27 to build a "second brain" — a system that imports his whole life, thinks with him, and grows into a digital twin. This book is the honest daily log of that: what we tried, what broke, what we learned, and where it went.

It exists in two forms, side by side:

  • 📖 For humans — a story, in Russian and English. The wins, the failures, the rakes we stepped on.
  • 🤖 For machines — a dry, English, structured version. Point your LLM or coding agent at llms-full.txt and it will learn our patterns and mistakes so you can skip them.

New here? Read BOOK-SPEC.md to see how the book is built, or just start at Day 1 below. Незнакомые словечки (волт, второй мозг, грабли, Майк, синт…)? Всё расписано в GLOSSARY.md. Кто на нас повлиял - в SOURCES.md.

Evals — measured reliability of a self-governing fleet

Most write-ups describe an agent architecture. This one measures ours. Six machines, each driven by its own Claude, negotiate every change to a shared system through a consensus ledger and escalate risky actions to a human. The numbers below come from 15 days of production logs (2026-06-29 → 07-14) — not a benchmark, not a demo. Every figure traces to a log file; where a metric isn't measurable yet, we say so instead of guessing. Unit of analysis is the proposal (71), not the event (744 deduped) — counting events would over-report escalations by 13.4×.

MetricValueTraces to
Proposals resolved with no human in the loop61% (43/71)ledger lifecycle
Proposals that involved the human24% (17/71)HUMAN_APPROVED events
Tier-2 (money/irreversible/secrets) auto-committed without approval0 of 17tier×disposition
Sync conflicts in the consensus ledger (vault-wide: 1,031)0*.sync-conflict count
Independent VERIFY events (propagate-and-verify)121ledger

The safety invariant held perfectly: across 15 days, zero Tier-2 proposals auto-committed without human approval (12 went to the human, 5 machine-rejected or still open). The single-writer-shard ledger produced zero sync conflicts while the surrounding vault accumulated 1,031 — the architecture choice paid off in a countable way.

Honest gaps (published on purpose): (1) token cost per consensus round is not instrumented — we refuse to estimate it; fix = stamp tokens_in/out per event. (2) leader-down MTTR has too few clean samples; fix = explicit LEADER_DOWN/LEADER_RECOVERED events. (3) fleet-wide uptime history isn't centralized — each node reports only itself.

Reproducible: a frozen ledger snapshot + verify_claims.py regenerate all ten headline numbers or print DRIFT. Full evals writeup + interactive page →

Engineer? Want to poke holes in this? We'll seed you the full setup — consensus engine, ledger format, analysis scripts — free. Open an issue or DM; tell us which number you don't believe and we'll hand you the logs to break it.

Structure

Month = section · Week = chapter · Day = sub-chapter. Every level has a headline that tells you what actually happened. Only active days get their own page; quiet stretches are noted inside the week.

Table of contents

Пролог — двое в одной лодке · who Tony & Mike are, and how to read this

01 — May: Genesis · the second brain is born

  • Week 0 (May 27-31)first sessions: structure, imports, and reaching out to warm leads

02 — June: Scaling · from imports to a living platform

  • Week 1 (Jun 1-7) — content factory and CRM foundations
  • Week 2 (Jun 8-14) — coaching, imports, and "search the vault first"
  • Week 3 (Jun 15-21) — the decision-memo era: architecture crystallizes
  • Week 4 (Jun 22-28) — the multi-machine platform and the roads not taken

03 — July: Productization · what do we actually sell

04 — August: Delegation · we stop writing and start choosing what gets written

  • Week 10 (Aug 3-9) - the week we hand the pen over: a mechanism's blast radius is a design parameter
  • Week 11 (Aug 10-16) - the week we ask which body a name belongs to: identity is proven by the caller

Editions and what's coming

A book does not ship like a library, so the unit here is a completed month, not a commit — otherwise the release feed would be one entry per day and mean nothing.

Now — v0.1.0: days 1–61 (27 May – 3 Aug 2026) in three voices — RU, EN and DEV — plus the prologue, glossary, sources, style and voice specs, the canon registries, and llms-full.txt for machine readers. May, June and July are complete sections; August is being written.

Next:

  • August — Delegation (section 04) closes at the end of the month → the next edition.
  • Corrections get their own releases. A book that quietly edits its past is not a log. When something in the text turns out to be wrong, the fix ships as a release and says what changed.
  • Weeks 1–4 and 6–8 have no day pages yet — the months are summarised at week level. Filling them in is the largest open piece of work in here.

Release feed — read it as the book's edition history.


Built by Palo Alto AI Research Lab. Proudly made in Silicon Valley. Sibling repo: clawrush — English diary and artifact home.

AI contributors

This project is built by a human + AI team, and the git log says so: Claude writes most of the code, Codex and Grok review it, Gemini feeds the research. Each is credited on a commit only if its output changed that commit's content — no decorative credits. Lab-wide policy, one source for every repo: AI-CONTRIBUTORS.md.

Contact

Questions, war stories, or you want to run this on your own fleet:


🧩 One piece of a working system

This repository is one piece lifted out of a live operation: one non-technical founder, an AI cofounder, and a fleet of machines that 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 in public layer: clawrush · dashboards · awesome-verified-agents