Best of 2026 (So Far)

These are the posts from 2026 I’d send to someone who asked what this blog is about. They cluster around four themes: agentic coding infrastructure, writing tools, long-running product research, and the plumbing that makes the rest possible.

OpenClaw: AI agent framework that runs on your own infra (April)

OpenClaw was the biggest project of the year. An open-source AI agent framework that runs on your own infrastructure — no cloud dependency, no per-token billing surprises, no vendor lock-in.

It uses background memory consolidation: between tasks, the system processes and compresses what it learned into durable memory, without blocking the next interaction. It forces workflow design rather than prompt chaining — the architecture is built around composable steps with explicit inputs, outputs, and approval gates. And it runs locally, which means sensitive data never leaves your network.

WordFloor (June)

After 13 years of running WriteDaily — a daily writing habit tool — I started something new. WordFloor is a private writing engine with a deliberately simple mechanic: set a word floor, hit the target, the square turns green. Delete everything afterwards — the green stays.

The insight behind it is that streaks are more sustainable when the content is disposable. WriteDaily proved that people will show up daily for years when the bar is low enough. WordFloor is the next iteration: practice without the archive, the streak without the content.

The waitlist is open.

WriteDaily, 13 years of user research (March)

Thirteen years of data from a writing tool nobody paid for. That’s an unusual dataset — most product research comes from active, paying users. WriteDaily’s users were neither: they showed up to write because the tool asked them to, not because they had a financial incentive.

The findings surprised me. Writers don’t want more features. They don’t want smarter analytics, better export options, or AI-assisted drafting. They want less friction between them and the blank page. Every feature request that survived triage was, at bottom, about removing an obstacle to starting.

Charts are in the post. The raw data is there too.

Blog refresh (June)

The least visible post of the year, and probably the one that took the most engineering time. I moved the blog to Astro, rebuilt the RSS feeds, and added a syndication pipeline that schedules posts to social channels at per-platform times with per-platform copy.

Most of this is invisible to readers, which is the point. A blog’s publishing machinery should be boring. The syndication pipeline uses authored bodies only — no runtime URL appending, no copy rewriting — and the whole thing runs on serverless functions with no persistent infrastructure.

Agentic coding infrastructure, writing tools, product research, and the plumbing that connects them. The individual posts have the detail — these are just the headlines.