Packaging one engineer's live opinions as an agent skill is a pattern worth stealing

What /kun does is straightforward to describe and genuinely odd to sit with: it packages a specific person's engineering judgment as an installable agent skill, then keeps that judgment current by scraping the person's public output every day.

The mechanism is worth understanding precisely. The skill file itself is intentionally thin - a loader, not a knowledge store. Every invocation fetches four markdown files directly from the repo over raw GitHub URLs: an entry point that tells the agent how to navigate the others, a tools index, an opinions map, and a voice guide. Session caching means repeated calls don't re-download, but the underlying files on main are updated daily by automation that reads the author's X posts, Substack, and YouTube. The skill you installed last month and the skill you run today are behaviorally different, because the knowledge base has been rewritten in the interim.

That daily automation is the genuinely interesting design choice. Most "persona" or "expert" prompts are frozen at authoring time and drift from reality immediately. This one treats the author's public statements as a continuous signal feed, merging new positions into existing files rather than appending indefinitely. The stated policy is to append only when something is truly new - which implies a compression step, not just accumulation.

The professional framing - L8 principal engineer at Meta, Microsoft, and Atlassian - is doing real work here. It's not decoration; it's the entire value proposition. You're not installing a generic "senior engineer" persona. You're installing a specific person's documented opinions about tools, workflows, and how to think through problems. The repo explicitly refuses pull requests because it is, literally, one person's knowledge base. That's a coherent position, not arrogance.

The practical questions this raises are real. What happens when the daily automation misrepresents a nuanced position by compressing it? How does the agent behave when OPINIONS.md contradicts the user's context? The README doesn't address either. The jsDelivr fallback for raw GitHub fetches is mentioned but not elaborated - it's a reliability hedge, not a guarantee.

For agent builders, the architectural pattern here is worth stealing regardless of whether you want Kun's opinions specifically: a thin skill wrapper that pulls live markdown from a version-controlled source, with a defined update cadence and a clear separation between the loader and the knowledge. That's a more honest model than baking a persona into a static system prompt and pretending it ages well.