Anton Dziatkovskii

Building an AI digital twin in public with Claude Code - English diary, longreads & reusable artifacts. RU version: Telegram @ClawRush.

Make Claude Code behave consistently across repos, sessions and humans: rules-as-files governance with precedence, declined-decisions journal, objection sparring. Sanitized from a live production system. Free, MIT.

A second brain for Claude Code and Codex that keeps receipts: skills as a plugin marketplace, opt-in persistent memory over MCP, vault templates and a CRM engine. Method, not data. MIT.

Run AI agents across machines without state drift: consensus protocol (propose/counter/accept/commit), dual-rail message bus, ACK discipline, self-healing sync. Formerly claude-consensus. Free, MIT.

LEASH-8: an 8-domain control model for AI agents with delegated authority. Scorecard, approval-design checklist, plan-vs-authorize pattern. Patterns from a live production agent operation, sanitized. Free, MIT.

CharmOS / C(H+A)RM — open-source framework to manage humans AND AI agents as first-class relationships, on a personal second brain, driven by the RDR loop (Recall to Deep Research). Manifest-first, Apache-2.0.

Catch fabricated RAG citations before they reach the user: a zero-token verbatim gate + a burden-of-proof judge, framework-agnostic.

Pre-registered A/B/C benchmark: Superpowers vs OpenSpec vs baseline, on real brownfield ops code. Methodology frozen before runs; results published either way.

Your LLM reviewer said APPROVE. Did it? A structured verdict contract: prompt rule + parser + exit-code gate in one stdlib file, with the 42 counterexamples that forced every line. Free, MIT.

Integrity scenarios for agent runtimes, distilled from real fleet incidents — deterministic fault-injection checks against real SDKs

Tools that produce evidence about what an AI agent actually did — gates, attestation, output verification, benchmarks. Inclusion bar: name the artifact a human can inspect later.

Keep Claude Code's always-loaded MEMORY.md from silently truncating: budget guard, orphan coverage, verbatim domain folding into hubs. Deterministic, no LLM for the mechanical parts.

One shared MCP daemon per machine instead of a stdio copy in every agent session: recipe, autostart templates for Windows/macOS/Linux, a watchdog that will not blind your live sessions, and the measurements to prove it

One persona, one frozen memory, N models: how much of an agent's character survives a model swap? Harness + blind multi-lens judge panel + cross-vendor rank control + contamination probe. Results included: 7 models, spread 4.80 -> 2.20.

Your scheduled job says exit 0 — prove it did the work. Three stdlib-only checks: output freshness, silent no-op detection, rollout proof by reading the fact back. Sanitized from a live agent fleet. Free, MIT.

Voice -> text -> your personal knowledge base. A primitive, not a platform: 4 small Python scripts turn voice notes into linked, tagged, searchable markdown.