/dev/michael

Code Mode is the idea that you stop describing tools to a model and start handing it a runtime instead.

I've been playing with browser use lately.

I'm building Mercury - my own take on a claw-like system.

Steering an agent is the act of reinforcing good behaviors and discouraging bad ones.

npm install -g napkin-ai pi install npm:napkin-ai Source: github.com/Michaelliv/napkin I've been avoiding this for a while now (years in LLM time).

pi install npm:pi-generative-ui Source: github.com/Michaelliv/pi-generative-ui The Discovery Anthropic announced generative UI for Claude a couple of hours ago.

Hello, World.

The backlog is dying and nobody is mourning it.

Agents know syntax.

There's a spectrum of what LLMs can build for people who don't code, and nobody's really mapped it.

Every agent harness starts with the same four tools: read, write, edit, bash.

Three months ago, personal agents weren't a category.

Your eval doesn't test what you think it tests.

A developer posted a 15-stage multi-agent pipeline that ships 2,800 lines a day through Claude Code.

Embed your docs.

Anthropic put 16 Claude agents on a shared Git repo and told them to write a C compiler in Rust.

Claude Code has a memory system that's not in the docs.

The terminal is where agents got serious.

Okay, I lied.

Claude Code hit $1B in run-rate revenue.

psst is designed so agents don't have to see secrets. psst SECRET -- curl ... injects the value at runtime.

The internet told me to add an llms.txt file.

Steve Yegge built Beads to give coding agents memory.

I can explain any feature in my codebases.

Claude Code stores everything locally.

A new spec dropped: Open Responses.

Watching Claude Code work is... text.

Skills are markdown files with optional packages attached.

psst started as one thing: let agents use secrets without seeing them. v0.2.0 makes it a proper secrets workflow. psst run The original pattern was psst SECRET -- command.

Anthropic uses bubblewrap for Claude Code, gVisor for Claude web.

I don't memorize command flags.

psst shipped.

I built a skill that makes Claude debate itself.

I have a confession.

Yesterday I wrote about context engineering needing an engine.

"Context engineering" is having a moment.

Your tools aren't capabilities you give the model.

A few weeks ago, Anthropic gave Claude filesystem access.

I wanted a way to share Claude Code sessions in blog posts.

I'll be writing a lot about LLMs and agentic systems here.

This is a meta post.