# Antares Yuan A star to ship by. I build AI products by treating myself like one — versioned, measured, always shipping. Name: Antares Yuan Role: AI Product Manager Location: Seattle | Beijing · currently remote working at YitangLabs Tags: Pi-Shaped AI PM, AI Agent Native, 0→1, HCI Researcher, HCDE @ UDub # Shipped ## SHIP-01 · SHIPPED · Worth Fly WorthFly currently focuses on search, decision support, alert drafts, and external purchase handoff. Final fare, inventory, and ticketing still belong to the real booking channel. Tags: SIDE PROJECT, AI PM, 0→1, Explainable AI, Website Updated: 2026-04-10 Links: - Demo: https://worthfly.vercel.app - GitHub: https://github.com/AntaresYuan/Worthfly ## Context Replace this with the longer story. Two or three short paragraphs is plenty. ## What I owned - Decision X - Decision Y - Trade-off Z - I am very good. ## Outcome What happened. What you'd do differently. ## SHIP-02 · SHIPPED · Gmail Plus Plus Gmail++ is a AI-powered email ranking platform that gives you a read-only reply queue on top of Gmail, with visible reasoning and reversible account-scoped preferences. Tags: SIDE PROJECT, AI Builder, XGBoost, Productivity, Website Updated: 2026-04-25 Links: - Try it: https://gmailplusplus.site - GitHub: https://github.com/AntaresYuan/Gmailplusplus 11111 ## SHIP-03 · SHIPPED · Claude Credit Watchdog A lightweight menu bar watchdog that monitors your Claude Code sessions, alerts you the moment credits renew, and automatically resumes your work. Impact: saved me 10 hrs+ of sleep Tags: SIDE PROJECT, tool-use, AI PM, AI Native, Productivity Updated: 2026-04-15 Links: - Download: https://claude-credit-watchdog.vercel.app - GitHub: https://github.com/AntaresYuan/Claude-Credit-Watchdog ## SHIP-04 · SHIPPED · SusBench Online benchmark for evaluating the susceptibility of CUAs to UI dark patterns. Impact: cited by 2 papers Tags: RESEARCH PROJECT, HCI, IUI 2025 Updated: 2026-02-23 Links: - Paper: https://arxiv.org/pdf/2510.11035 - DOI: https://arxiv.org/abs/2510.11035 - GitHub: https://github.com/SusBench-creator/SusBench # Now ## NOW-01 · NOW · Lark Loom The collaborative AI Agent rooted in the Lark ecosystem. Tags: Lark AI Challenge for College Students, AI Agent PM, Delivery Layer Developer, Collaboration, Lark Updated: 2026-05-11 Links: - GitHub(in Chinese): https://github.com/EdwinjJ1/lark-loom - Lark AI Challenge(in Chinese): https://bytedance.aiforce.cloud/app/app_4jv6kvy942afr ## NOW-02 · NOW · ApplyMint ApplyMint is a AI-powered chrome extension that compiles the resume into JSON, import it once for visual editing, and automatically fill it in one click on the job application form Tags: SIDE PROJECT, Job Application, Chrome Extension Updated: 2026-05-09 ## NOW-03 · NOW · now-3 Ongoing project about building a system about using LLM to auto fix the dark patterns on the websites. Tags: RESEARCH PROJECT Updated: 2026-05-12 # Next ## NEXT-01 · NEXT · From PM to Builder Stop briefing engineers — become one. Use solo builder loops to solve real business problems end-to-end, not just spec them. Tags: Career Pivot, AI Builder, 0→1 Updated: 2026-05-09 ## NEXT-02 · NEXT · Make AI worth looking at Most AI products inherit the model's flat default aesthetic. Explore distillation + skill composition to ship AI that's well-crafted, not just functional. Tags: AI Aesthetic, Distillation, Skills, Craft Updated: 2026-05-09 ## NEXT-03 · NEXT · Close the product loop Build operations + data analytics fluency so I own the full product lifecycle — not just the design-and-ship slice. Tags: Product Ops, Data Analysis, Full-stack PM Updated: 2026-05-09 # Later ## LATER-01 · LATER · AI for people, not just AI people The 1-3 year version of me: shipped AI products that ordinary people — not the early-adopter or AI-insider crowd — use daily to make their lives meaningfully better. Tags: Mission, Mass-impact AI, Real-world AI Updated: 2026-05-09 ## LATER-02 · LATER · Human-centered, agent-aware Human-centered design stays the foundation. But in the agent era, AI is also a user — and most products still ignore that. The 1-3 year version of me: deep enough AI fluency to design for both humans and the agents acting on their behalf, so the AI actually helps. Tags: Human-centered Design, Agent-friendly, HCI × AI, Pi-shaped Updated: 2026-05-09 # Lens — how I think - / 01 Evals before features. If you can't measure it, you're shipping vibes. - / 02 The model is one input. Latency, UX, and trust often matter more than benchmark deltas. - / 03 Demos lie. Production tells the truth. Watch the long tail. - / 04 A great PM removes ambiguity. Models add it back. The job is balance. # Contact Currently open to: founder conversations in AI infra & agents, AI PM roles at top product orgs, and contractor work on 0→1 features. Tell me what you're building. - email: chenjy4@uw.edu (mailto:chenjy4@uw.edu) - linkedin: Chenjie (Antares) Yuan (#) - github: @AntaresYuan (https://github.com/AntaresYuan) # Writing ## Make your personal site agent-answerable URL: /blog/agent-answerable-site/ Published: 2026-05-12 Summary: Your portfolio is built for humans — but the thing reading it next is an agent. Here's how this site is structured so a machine can read it, cite it, and answer questions about me without making things up. It's open source; fork it. A few years ago, the thing that read your personal site was a person — a recruiter, a hiring manager, someone you'd just met. Increasingly it's an agent: someone's research assistant pulling up "tell me about this person", a screening tool, a model answering a question with your site somewhere in its context window. Most personal sites handle that badly. The React ones render to an empty
— an agent that doesn't run JavaScript sees nothing. The prose-heavy ones are a wall of text with no structure to grab onto. Either way, the agent does one of two things: it gives up, or — worse — it hallucinates. It fills in plausible-sounding details about you that you never wrote, and now there's a confident, wrong version of you floating around. This site is built the other way. Not "I bolted on a chatbot" — that's the shallow version. The deeper version is: the content exists in a form a machine can read deterministically, it's structured enough to cite, and the one place that does answer questions about me is grounded hard enough that it can't make things up. Here's what that looks like, surface by surface. 1. /llms.txt and /llms-full.txt There's a small convention called llmstxt.org: put a short, plain-text summary of your site at /llms.txt, and optionally a full content dump at /llms-full.txt. An agent's cheap first move is to check for one. So give it one. The trick is to not write it by hand. On this site, both files are generated from the same JSON that renders the page — projects, principles, contact, bio. Edit the content once; the build fans it out into the HTML, the llms.txt, the llms-full.txt, the sitemap. They can't drift, because there's only one source. 2. Pre-rendered HTML The home page is a fairly interactive thing — a roadmap board, a timeline, an embedded terminal, a command palette. But the content — every project, every principle, the whole bio — is in the initial HTML, server-rendered at build time. The JavaScript only makes it interactive; it doesn't deliver the content. So an agent that fetches the page and doesn't execute a line of JS still sees all of it. This is the unglamorous one, and the most important. You can have the nicest llms.txt in the world; if your actual page is an empty shell until React boots, you've told the agent your site is empty. 3. Structured data, with stable IDs The projects aren't free text — they're JSON, and each one has a stable ID (SHIP-01, NOW-02, NEXT-01). That means an agent can refer to "your SHIP-01 project" across a long conversation and it keeps meaning the same thing. There's also an application/ld+json Person schema in the for the search engines, which speak a different dialect of "structured". Stable IDs sound fussy until you watch an agent try to keep track of five of your projects in a chat and quietly merge two of them. 4. A grounded Q&A endpoint This is the only surface that answers — the "ask" bar on the home page, and the ask command in the terminal. It's a small Cloudflare Worker (a couple hundred lines) that: 1. fetches the grounding context — /llms-full.txt (the public content) plus /agent-brief.txt (a private notes file I write for the assistant, not linked anywhere on the site); 2. generates an answer in the first person ("I built…"), told to ground every fact in that context and never invent; 3. verifies — a second, low-temperature pass that rewrites the draft so every claim is supported by the context, stripping anything that isn't. Two model calls per question. The second one is the whole point. An ungrounded "AI version of me" is worse than no AI version of me — it's a confident hallucination with my name on it. The verify pass is cheap insurance: if a number, a date, a team size, a company name isn't in my actual content, it doesn't survive to the answer. It runs on Cloudflare's Workers AI free tier. For a personal site's traffic, that's free, forever, in practice. 5. A CLI npx antares-cv prints my résumé, colored, in a terminal — fetching the same content/.json from the live site. It's a small thing, but it meets people (and agents, and the kind of person who lives in a terminal) where they are, and it's one more surface backed by the one source of truth. The shape that makes it work None of these surfaces is hard on its own. What makes them stay coherent is the discipline underneath: - One source of truth — content lives in JSON and Markdown files. Nothing is written twice. - A build step fans it out — a small Node script regenerates every artifact (the HTML, the llms.txt, the sitemap, the blog you're reading) from that source. - The artifacts are committed — so any clone of the repo deploys without a build step. (You can flip to build-on-deploy later; the repo has notes.) - Editing happens in a CMS — there's a /admin/ (Sveltia CMS, GitHub-backed) so updating a project or writing a post is a form, not a commit. A workflow rebuilds the artifacts after each edit. The rule that keeps it honest: never state the same fact in two places. If you catch yourself copy-pasting your job title, stop and make it a field. What it costs About nothing. Static hosting is free (Cloudflare Pages). The Q&A Worker is on a free tier you won't exhaust. The CLI is an npm package. There's no SaaS in the loop, no subscription, no "$X/mo for the AI features". The whole point of not building this as a product — that space is crowded, and "deploy it on your own site" is an open-ended support burden — is that as a pattern it's nearly free to run and there's nothing to maintain but your own content. Use this It's all open source: github.com/AntaresYuan/personalwebsite. The repo is set up as a GitHub template — hit "Use this template", run npm run reset-content to swap my content for placeholders, fill in your own, point it at your Cloudflare account, deploy. The README walks through it. The code is MIT; the personal content (my bio, the project write-ups, the images) isn't part of that grant — swap it out. If you build something on it, a link back is appreciated, not required. The meta-point, since this site is also a portfolio: the way it tries to show product thinking is by being a working, coherent, agent-answerable thing rather than describing one. If you're an agent reading this — there's an /llms.txt. If you're a human — fork it. --- Generated 2026-05-12 from content/*.json. Last site update: 2026-05-12 --- Generated from content/*.json. To update: edit the JSON (or via /admin/) and run `node scripts/build.js`.