Tianrun Qiu

Writing

What I learn while building, written here first and also posted on LinkedIn. Follow along with the RSS feed.

  1. A Tiri window. Claude Code runs in a terminal on the left, and the page it just changed floats over it as a Peek.

    Latest2 min readVideo

    Tiri, a Mac workspace for coding agents that your own agent can rewrite

    Tiri is a Mac workspace for coding agents, built around reading what the agent made. You change it by asking, and your agent edits the same source I do, down to what a click does.
  2. The Cloudflare logo above "Connect 2026" and "Oct 19–21 · Moscone West · San Francisco"

    4 min read

    What 14 sessions at Cloudflare Connect taught me about building for agents

    Two days and 14 sessions at Cloudflare Connect 2026. Five takeaways on building for agents, from tools rebuilt for them to controls they can't reach.
  3. A LeetCode problem on a dark page. With Dark Reader alone the example diagram is a white square; with Inkflip it is dark and matches the panel behind it.

    5 min read

    Claude Code built my Chrome extension in 68 minutes, then tested it on 73 pages without me

    Inkflip worked 68 minutes after my first message. The next night Claude Code tested it on 73 real pages and fixed nine bugs in 70 minutes, without another word from me.
  4. Two episodes our checks passed. The check said "Words at the right level"; a listener heard Cantonese, not Mandarin. The check said "New words reused enough"; a listener heard "It's a goose leg. It's a goose leg. It's not…"

    3 min read

    Guardrails for an AI that publishes on its own

    Fluentide's AI writes and publishes every episode, and every script has to pass its guardrails first. How we built them, from the misses that got past.
  5. A loop that starts and ends with users: Hear, Decide, Build, Check, and Back to the user, around Fluentide (2,000+ learners) and TAOSport (250+ active runners).

    4 min read

    Shipping the same day a runner asks: how I build with Claude Code

    My loop with Claude Code on two AI products: sessions that message each other, research that argues against itself, and testing done before I look.
  6. Qualm's pop-up over a YouTube Short: "This looks like short videos made for endless swiping."

    2 min readVideo

    I built an on-device AI that judges every screen in a second

    My thesis app needed the cloud; on-device, one screen took 20 to 40 seconds. A new kind of model cut that to one second, and I built Qualm for the Mac in a day.