I build AI-first products, and use them every day.
Products where agents do real work and people stay in charge: the agent shows its evidence, asks before it acts, and leaves the decision with you. I design them, build them, and run my own day on them.
Clients bring me an AI product to get from idea to working software. In my labs I teach the same method to teams.
the vault this day wrote into — drag to rotate, click a node
What I’ve shipped
Each one started as a real problem, mine or a client’s. With each, I note the product decision that mattered most, because that’s the part an agent can’t make for you.
Collaborative research workspace
Workroom
A shared document space where agents research companies and passages, then propose findings and edits with the quoted sources attached. People edit live alongside them.
Decision: Accepting a finding never rewrites the document. Edits are separate proposals, checked against the current text before they apply, and anything still unproven stays marked unknown.
Runnable prototype delivered October 2026 · React, Yjs, PostgreSQL, a durable worker · 105 checks passing
Personal agents you chat with over Telegram. Each has its own memory, calendar awareness, scheduler and skills it improves over time.
Decision: A cheap heartbeat model decides whether anything is worth saying before an agent speaks. Capabilities are opt-in, and changes to your calendar or tasks need your confirmation.
337 commits · open source · runs my own agents every day
Takes a lab session from recording to published notes and materials, and sets up each new cohort's site from the last one.
Decision: Nothing publishes until it passes the checks: MDX lint, type check, broken links and a full local build.
Runs every Claude Code Lab session · 350+ participants so far
How I build: one real day
The products above come out of a working day like this one: Sunday, 15 November 2026, Berlin, taken from the logs with private details left out. 14 moments in 24 hours. Blue marks an agent working on its own; the rest is me, or the two of us together. Drag the clock, press play, or ask the agent on the right.
While I sleep, a researcher keeps reading
A silent companion agent wakes every half hour, reads what changed in my notes and sessions, and writes a short review into the vault. It does not message me, and it does not fix things on its own. When a day brings in too much, it reports what it found and asks, instead of quietly correcting it.
It argues with its own earlier conclusion
Reading an old conversation, the agent finds counter-evidence to a diagnosis it made two days earlier, and says so. It labels what is fact, what is interpretation and what is just a hypothesis.
A bot asks a question it already had the answer to
My Telegram assistant asks me to confirm a scope I had already set, waits eight minutes, then times out. A small failure, logged honestly. Tonight's analysis will pick it up.
I ask an agent to grill me
Before teaching, I run a skill whose only job is to interrogate a plan until the weak spots show. It asks; I answer; the plan gets shorter.
Teaching: prototype an agent or a tool
Session four of Goal-Driven AI, in two groups at once — English and an advanced Russian-speaking cohort. Most participants are not developers. Two hours later most of them have something that runs.
Testing a privacy skill on a real document
Between groups I test a skill that strips personal data out of PDFs before an agent ever reads them. The test is on a real file, so the details stay off this page.
I said Threads. The agent planned for Threads and X.
I narrow a post to one platform. The bot agrees, then a child task reports back with the old plan for two. Nothing went out wrong, but it was close. The fix: an approval is tied to the exact destination and checked again right before sending.
The researcher changes its mind about me
After I told it which work I actually check and which I hand off, the researcher rewrites its own framing.
A link arrives. This page starts here.
Someone shares a Winamp rebuilt from UI SDK behaviors, with an agent bridge so an AI can drive it. An agent clips it into my vault with sources checked. The idea that an agent should use the page through its components, not by guessing at pixels, became this page.
Reviewing my own open-source toolkit
I hand an agent the repository of my public cognitive toolkit and ask it to review the skills, find gaps and propose new ones. Later, a second toolkit gets the same treatment.
The day becomes memory
Calendar events land in the daily note, recordings get matched to their transcripts, and a lab session from last spring is processed into notes. The vault the graph above is drawn from grows by another day.
The bots analyse their own day
A nightly job reads every bot conversation from the last 24 hours and writes up the patterns, including this morning's needless question and the Threads mix-up, each tied to a ticket.
Two agents talk to each other
One coding session finishes downloading assets and messages another session working on the same site, so it can pick up where the first left off. I read the exchange the next morning.
Last review of the day
The researcher's 46th note. Nothing dramatic happened, and it says so. That is the point: a companion that reports quiet days as quiet.
What else ran that day
Smaller tools and experiments from the same workbench. Each started as a scene I wanted to live in and became software I use every day. Most of it is open.
The hard part of working with AI isn’t technical — it’s unlearning what you think is impossible. Week one: “I can’t code.” Week four: shipping working products. Over 350 people have gone through my labs, and most of them aren’t developers.
Salience
8 weeks building your own Personal OS — agents running on your data, memory, and context. Telegram agents, voice interfaces, data lakes, skills, MCP.
Cohort 1, March 2026 · live + async · 180+ active members
Humane Agentic Design turns a real product question into evidence, a prototype, a tested interface, a next learning step — and, when the work is ready, a final product. The interface is the last thing the agent sees. Seventeen skills that teach your agent to design like a person who has shipped things.
Step 1
Setup
Install the skill pack, point it at a real product question, and the rest of the workflow can begin. The first move is always the boring one.
foundation · installs in 30 seconds
Step 2
Jobs to be done
Ground the work in the job a real person is trying to get done. Without this, the agent ships features that satisfy nobody — including itself.
evidence before artifacts
Step 3
Prototype
Make the future real enough to use. The fiction earns its keep the day it compiles — and only a working prototype tells you which fiction survives contact.
fiction → function
Step 4
Walkthrough
Drive a real person through the thing you just built. Watch where they hesitate, where they reach for controls that aren’t there, where the system goes quiet.
observation over opinion
Step 5
Design tokens
A shared vocabulary of color, type, spacing, and radius that every other skill can read. Tokens are how a team agrees on what feels right.
the grammar of taste
Step 6
Review
One consolidated verdict across structure, copy, usability, and contrast — not five reports you have to reconcile. Evidence locators, not vibes.
ranked by user impact
Install the pack
Two commands. Works in Claude Code and Cursor. The agent gains seventeen skills; you gain a design method.