This is the short path through the work: one story, four stops, and a clear rule for what stays public. It does not try to list every repo; it shows the operating loop a visitor can inspect.
Public rule
Keep what strengthens the public operating-system thesis.
core stops kept
public narrative
extra demo branches
Start from repeated friction in building, reviewing, writing, or using AI tools, not from abstract novelty.
Turn the pain into a repo, route, scorecard, or repeatable workflow that can be inspected outside the chat.
Expose only the safe, useful part: a page, demo path, README, report, or technical write-up.
Useful output becomes a rule, page, note, or tool improvement so the next 30-minute loop starts higher.
Next step
The demo explains the loop. The build log shows what shipped recently, why it matters, how to inspect it, and what stays private.
Suggested 4-minute walkthrough
Everything else should support these stops, not compete with them. Supporting repos remain in the proof chain and GitHub profile, but this page should stay sharp.
A file-first operating layer that turns my knowledge base into a working digital twin for research, writing, site updates, and continuous refinement.
Demo line: The thesis: files, agents, and write-back loops can form a compounding personal operating system.
A trace-first standard and CLI for judging whether an AI agent is worth more token budget, permissions, and autonomy in my real workflows.
Demo line: The filter: agent work is judged by artifacts, tool follow-through, verification, and review burden.
A proxy server that exposes OpenAI/Gemini/Claude/Codex-compatible APIs for CLI and coding tools, with OAuth login, provider routing, and multi-account load balancing.
Demo line: The infrastructure: CLI/OAuth/model-provider access becomes a routable API surface for coding agents.
My public narrative system: a Next.js site, MDX writing archive, and AI-assisted publishing workflow designed to compound proof over time.
Demo line: The public channel: useful learning becomes writing, demos, and proof that a stranger can inspect quickly.
What gets cut
Manage Up, Claude Code Sourcemap, Knowledge Harness, and older experiments still matter as proof. They just should not all fight for first-screen attention unless they advance the same core loop.
Keep / cut scale