Writing

A practical walkthrough of OpenClaw's five memory mechanisms, how they work together, and why I still prefer a local-first approach to agent memory.

AutoHarness and Anthropic point to the same lesson: agent performance depends less on prompt cleverness than on the harness that constrains, evaluates, and carries work forward.

If stable AI performance depends on reusable skills, then skills themselves need a review method before they can be trusted.

How I turned Codex into a narrow repo-specific experiment loop, with a generated program.md and a Kaggle competition repo as the proving ground.

From token growth to autoresearch and OpenClaw, the defining shift in this AI cycle is not just smarter models, but systems that keep running, spending resources, and moving themselves forward.

A practical OpenClaw upgrade checklist that starts with backups, verifies the gateway and plugins first, avoids a bad release, and documents how I recovered the WeChat plugin after the SDK migration.

How I turned OpenClaw from one main agent into a small multi-agent team, starting with one coding specialist and a plan for research, operations, and backlog sharing.

What I actually did to get OpenClaw working from a clean reset, including Slack, Tavily, Lossless Claw, QMD, and the parts that still failed.

A simple way I use to see whether a new AI tool is actually useful in real tasks.