Things I'm building, testing, and learning from.

Agent Native Composer treats a folder as the durable workspace for an AI agent. The folder holds the agent's identity, instructions, skills, memory, current task, and runtime state. The document in the focused editor pane becomes that agent's active document.

Opening another document changes the focus within the same workspace. Switching to another folder brings forward its own agent without interrupting background work in the first one. Each running task keeps the workspace, document focus, permissions, and context it started with.

The editor stores work-relevant memory in readable project files: instructions, a compact context snapshot, an activity log, the current task, and per-document checkpoints. Provider threads can be replaced when needed because the folder, not a chat session, carries the durable record of the work.

The first runtime uses the locally installed Codex app server and its existing ChatGPT sign-in. OpenKnowledge provides the Markdown editor, file navigation, collaboration, history, and desktop foundation. Folder-aware agent behavior sits above that foundation as a focused system for writing and building with an agent that remains connected to the work.

Learning Capture is a reusable Codex skill for turning a YouTube link into a durable learning report. It fetches metadata and timestamped captions, asks the agent to read the transcript before judging it, then saves a Markdown report with the transferable process, named tools, missing details, critique, and practical experiments worth trying.

The important shift is from a personal OpenClaw automation to a portable skill. The workflow now lives as SKILL.md plus a tested transcript-fetching script, report templates, and Codex/OpenAI interface metadata. That makes it easier to install in a new workspace, run from a pasted YouTube URL, and improve through normal repo review.

The skill also gives the agent a stronger evidence standard. Claims are labeled as demonstrated, explained, asserted, or opinion, and absent details stay marked as undisclosed instead of being filled in. That keeps the reports useful for learning without turning every video into a vague summary.

The visual system guide started as an AI-led brand interview, then moved into visual-direction exploration, a shareable HTML guide, and a website-specific homepage concept. After the guide existed visually, I generated and refined docs/design.md so coding agents could use the brand system while building websites, apps, newsletters, and other digital assets.

The project is less a reusable toolkit item and more a case study in turning personal positioning into an operating design system.

Caddy is a personal command-center project for OpenClaw-assisted work. The nearby dashboard app uses Next.js App Router, Tailwind, shadcn-style components, and local data/API routes to bring together AI learning logs, sports calendars, quick capture, and side-project metrics.

The goal is to make the assistant workflow more visible and operational: a place to see what is in motion, capture new inputs quickly, and keep personal systems close to the work.

I wanted Codex to control the Sonos system from the same workspace where I was already working. Not through OpenClaw, and not through a cloud API. Just local network discovery, status checks, play and pause, volume changes, group status, and favorites.

I one-shotted the first version with /goal. The brief was simple: make a small tool that Codex could use safely from the terminal, then write the agent instructions that translate ordinary requests like "turn down the office" into concrete commands.

The useful part was that the project was made for Codex and with Codex. It did not need a big app surface. It needed a clear command line, JSON output, conservative defaults, and instructions that made passive checks the default until I explicitly asked it to change playback.