At first I thought this was just grunt work. Create some folders, fill in some templates. But the most interesting part turned out to be the question I asked at the end myself: "what's actually different from the skills I already had?" Running mail-triage didn't feel any different at all, because that skill never even touches the new context files, it's completely self-contained logic. The real difference isn't in running a specific task, it's in casually asking something like "how should I plan this week" later on. it'll already know the order is, already know my app launches next weekend, without me having to re-explain the background every time. The interview part was interesting too. Answering the same set of questions again (team, priorities, communication style). I noticed I actually answered more specifically this time than last time (the grill-me session on Sep 9). One more small discovery: running mail-triage, I found that the last round (Aug 28) had already processed over 8,000 emails. this time was just the incremental 536. Looking back, I realized this system has actually been quietly running for almost two weeks, I just hadn't noticed "oh, I've been using this the whole time.

What you scheduled: A daily job digest that runs every morning at 7:30. It searches for full-time remote roles, filters them against my own search criteria, and only surfaces genuinely new matches. Scheduled task or loop, and why: Scheduled Task, not Loop. This is an ongoing daily habit with no end date, not a short-lived window of help. Loop is session-bound with a hard 3-day expiry and no catch-up if you miss a run. One surprise: Running it twice back-to-back returned the exact same results both times, which taught me the platforms don't push matches in real time. Zero new results wasn't a bug, it just meant "check back next cycle."

Live URL: https://gankan-test-website.vercel.app This is the landing page for GanKan (敢看), an app I'm building. It turns every expense into "how many hours of work that cost you" instead of just tracking numbers. It's Chinese-only for now, English version coming later. The app itself hasn't launched yet, so this page is a preview of where the design is at. The hack that made the biggest difference: cloning a reference site by pulling the real source code instead of guessing from a screenshot. That's how I caught a fade-up scroll animation that turned out to be dead code in their production CSS. it was sitting in the stylesheet but not actually running live. Would've copied an effect that doesn't even work if I'd just eyeballed it from the screenshot. Before/after: early draft was a flat light-purple layout, everything stacked in equal-width blocks, no motion. Current version moved to a dark immersive hero with a gradient wave, a live percentage ring, and alternating dark/light sections for rhythm. Pretty happy with this first version, but there are still real gaps in the logic/flow I want to fix. I will keep iterating on it.

What you deployed content material inbox in notion What it does A weekly automation that pulls the latest videos from 3 YouTube creators I follow, dedupes against what's already logged, and drops title/link/summary into a Notion database, building a raw content-idea inbox for my own social media (Xiaohongshu/Douyin) without me manually browsing YouTube every week. One thing that broke The summary extractor was grabbing the wrong line from video descriptions, for creators whose description starts with a promo/CTA line ("Subscribe for more...") or a timestamp, it picked that instead of real content, so multiple videos ended up with the exact same summary. Claude Code added pattern-matching to skip promo/timestamp lines and fall back to the video title when no real description line exists — no LLM involved, just smarter string parsing.

My Skill: project-briefing What it does:· I often deal with messy folders (code projects, messy notes, or desktop clutter). I built a skill called project-briefing that scans a directory structure, skips noise directories, flags sensitive files by name only (never touching .env or keys), and writes a condensed _briefing.md summary directly inside the folder with an overview, topic groupings, recent changes, and unfinished files. How to trigger it: - I can trigger it using natural language like "summarize this folder", or via the /project-briefing command followed by the path. - The bug and the optimization I made: I tested it on a real course folder containing just 3 files, the agent successfully read two PDFs, but hit an Outlook .msg email and silently wrote: "couldn't read this format, skipped." It turned out that skipped email was the most important file in the folder—an internal message revealing that the PDFs were recycled from an old year's template, which finally explained a date that made no sense otherwise. Watching the agent quietly give up on the one file that actually mattered taught me a vital lesson. I added a new rule: for common non-native formats (.msg, .eml, .docx, .pptx), it must actually try to extract the content first and only report "unreadable" if it genuinely fails.