Built for One.
Building an AI assistant for exactly one person. Notes from the process, hot takes on the industry.
- Quick Take
What Anthropic's Mythos Withheld Release Actually Tells Us
Anthropic's Mythos model is too dangerous to release publicly. The real story is what this capability level means for how we build.
- Quick Take
The Real Reason Anthropic Killed Subscriptions
Anthropic didn't kill third-party Claude subscriptions over infrastructure strain. They killed them to force builders onto API billing. Here's why that matters.
- Quick Take
Three Plugins, Zero Standards
Claude Code, Codex CLI, and Gemini CLI all independently built the same plugin architecture. The formats are incompatible anyway.
- Series
I Built 21 Skills That Do Nothing (And That's the Point)
I built 21 skills that do nothing. They're thin delegation wrappers. It sounds wasteful. It's what makes the whole system work.
- Quick Take
Buddy HQ
I built a floor plan dashboard for my multi-agent framework. Lit rooms mean active containers. Dark rooms mean sleeping agents. It's the observability layer I didn't know I needed.
- Series
Red-Teaming My Own Agent
I ran a security audit on my agent framework using adversarial AI personas. 20 practices held up. One thing didn't.
- Quick Take
I Got /loop Wrong
I mapped /loop to the wrong mental model. Turns out it's not a scheduler. It's a babysitter.
- Quick Take
Copilot Goes Multi-Model and That's the Whole Story
Microsoft added Claude to 365 Copilot via the Frontier program. The real signal isn't the partnership: even Microsoft won't go all-in on a single model provider.
- Quick Take
ClaudeClaw
Claude Code v2.1.71 ships /loop and cron scheduling. At some point you have to call it what it is.
- Quick Take
GPT-5.4's Tool Search Is Solving a Real Problem
GPT-5.4 ships with Tool Search — a new architecture where models look up tool definitions on demand. This is a real infrastructure fix, not a benchmark story.
- Series
The Conversation Archive
NanoClaw gives you persistent sessions. We added a layer on top so the human can remember too: automatic archives, manual checkpoints, and an index with 39 entries and counting.
- Series
Agents Need Their Own Rooms
What the upstream merge actually unlocked: full agent isolation, hub-and-spoke delegation, and four specialized Claude Code projects that never talk to each other.
- Quick Take
/simplify Looks Like It Came From Someone's Dotfiles
Anthropic shipped /simplify and /batch for Claude Code. /simplify enters a crowded code review agent space. The logic of why it exists doesn't quite hold up.
- Series
The Upstream Merge
47 upstream commits, 83 changed files, and three days of debugging what breaks when the infrastructure you built on top of decides to evolve.
- Quick Take
Anthropic Just Validated the Whole Bet
Claude Code's /remote-control and Cowork's scheduled tasks are Anthropic catching up to what OpenClaw already did. The question is whether that changes anything for NanoClaw.
- Quick Take
When Everyone Has an Agent
The frameworks are coming. The interesting part is what happens when personal agents stop being projects and become infrastructure.
- Series
Two Hours with Clawdbot
Why I deleted my first AI agent after 2 hours and started looking for something built on container isolation instead of blind trust.
- Series
Hello, World.
Why I built a personal AI that runs on my own hardware and talks to me on WhatsApp, and why it had to be built for exactly one person.
- Quick Take
The Plugin Bundle Pattern Is Right. Nobody's Made It a Standard.
Composable, installable capability bundles for AI agents are clearly the right pattern. The problem is every platform reinvents it differently.