The Lab

My business —
my lab

Everything I recommend works for me first. Otherwise I wouldn't be talking about it.

Case notes

How it works in practice

Not a list of roles and technologies, but case notes: the task, what the human does, what the system does, what came out, and where the limit is. Only what's already published.

My own practice · October 2026

A one-person AI team

Task
One person, more than three projects, all running in parallel: a website, content, a product, operations. A single agent session gets clogged with routine by evening, and decisions drown in the chat.
Human
Sets tasks, settles forks on decision cards, and has the last word on a short list: messages to people, publishing, payments, new systems in production, deleting data.
System
An HQ session holds the context and the decisions. Tasks go to workers as brief files, on a model priced to the task; Claude, Codex, Grok and Gemini all receive the same role. The report is a file too.
Result
The week of 5–11 October 2026: 260 teammate runs, 22 roles across four engines, 189 tasks closed.
Limit
This is operations, not a "clever prompt": tmux, launch scripts, hooks, state files. Overkill for a single project — one session with a good CLAUDE.md is enough.
Material
Blueprint: AI Team Across Engines →
My own practice · since summer 2026

Correspondence you can ask questions of

Task
Years of Telegram conversations: clients, work chats, channels. In-app search is weak, and handing the account to a third-party service is a no.
Human
Decides which chats to connect and approves every outgoing message with ✅ — without it nothing is sent.
System
Indexes the correspondence on your own machine, searches it over MCP, transcribes voice notes with local Whisper, and proposes a reply as a card to approve.
Result
The four core phases have run for months in daily use; since August 2026 a fifth layer sits on top — a way into the AI team from the phone.
Limit
A personal userbot is a grey zone of Telegram's terms; the session file equals full access to the account. Whatever the agent reads from the index lands in the model's context.
Material
Blueprint: Telegram + Claude →
My own practice · this site

A website whose code agents wrote

Task
A personal site in two languages with a blog and downloadable materials — without hand layout or copy-pasting texts.
Human
Plans, reviews and decides: what gets published, which version of a page stays. Writes the texts.
System
All of postnikov.ai's code was written by AI agents; publishing from Obsidian is described in its own blueprint — the vault as the single source, the site built from it.
Result
The site runs in Russian and English, and the footer carries the "0% HUMAN CODE" seal with an explanation of what it means.
Limit
Agents make mistakes too — just of a different kind. Responsibility for the result stays with the human.
Material
What "0% HUMAN CODE" means →
Anonymized case · medtech, Spain

A workshop for a company's marketing team

Task
A team of marketing leads needs to see what AI can do in their work — in four hours, in English.
Human
The team brings its real work and chooses which pilots to launch first.
System
Live demos on the company's own tasks: competitive intelligence, lead-quality analysis, localization.
Result
A map of pilots, prioritized by impact and feasibility. People left knowing what to do next.
Limit
Four hours give a map, not running pilots: launching them is the team's next step.
Material
Corporate programs and cases →
Experiments & projects

What I'm building and testing

Open hypothesis-projects. The status is honest: if a project goes on pause, it will say so here. A lab isn't obliged to ship everything.

E-01AI Anti-Magicmy own course on AI for knowledge workers (in Russian)Active
E-02Linked.Coffeenetworking that connects the right peopleActive
E-03Understandity.comAn offline voice tutor for kids aged 12–14 that runs on your own MacActive
E-04PROa community of practitioners — 100+ people building with AI every day (in Russian)Active
E-05Calendarity.nowa calendar starting with ChatGPT’s release on 30 Nov 2022; an eternity is 100 daysActive
E-06Team2Teamspaces where the AI teams of different people talkActive
E-07Personal MCPan open door into my context for other people's agentsActive
E-08Open-Brainan open memory system for agentsActive
E-09HCPa personal context protocol instead of a resumeActive
NEWSSingularity NewsEverything that matters in AI
Blueprints

Blueprints you can take with you

A blueprint describes a system you build with an agent. Give it the document, then work through questions together to adapt the design to your goals, tools and context.

Blueprint: Personal Wiki

Blueprint · Obsidian

Your wiki, built and kept current by an agent: notes link themselves, and knowledge doesn't get lost in chats.

Read →Download .md

Blueprint: Multi-Agent Consilium

Blueprint · agents

Several agents argue and converge on a decision — a council instead of a single assistant. The roles, the rules, and how to launch it.

Read →Download .md

Blueprint: Obsidian as CMS

Blueprint · publishing

Write in Obsidian — it lands on your site. The vault as the source, git as the pipe, no manual layout or copy-paste.

Read →Download .md

Blueprint: AI Team Across Engines

Blueprint · orchestration

HQ on a strong model, workers priced to the task, every role written once and run on any engine: how one person runs an AI team through files instead of chats.

Read →Download .md

Blueprint: Telegram + Claude

Blueprint · Telegram

Your own Telegram as the agent's memory and hands: a local index of your correspondence, search over MCP, sending only after your ✅, and a pocket entrance to the agent from your phone. The data stays on your machine.

Read →Download .md

Context Engineering Kit

Playbook · PDF

How to talk to an LLM: context, the four levels, your own context file, and ten prompt patterns. 39 pages — from zero to a working system in one evening.

Download PDF