Project Log · Starting Xreader

Start by asking why

Whenever I do something, I want to think one question through clearly first: why am I doing this at all?

Reviewing this year’s projects

Lately I’ve been reflecting on the projects I took on this year, from the start of the year until now — you can see them on my GitHub.

WYRD Nova. It actually came from Chunjun’s idea: interactive novels. Products like this already exist in abundance on Xiaohongshu. I later brought in Siyao, hoping she would handle distribution and operations, but I found her energy wasn’t in it — it felt forced. And the direction wasn’t one I liked anyway: I tried interactive novels myself, and they’re really games. I did put up an initial product framework, but the biggest problem with this project was that the generated content wasn’t good enough, not compelling enough — maybe I just hadn’t tuned it well. With LLM-driven dynamic generation, the barrier is fundamentally low and the market is already crowded. The deeper reason: it’s not a direction I like, nor one I understand deeply or am good at.

Mindset Bridge. Essentially a product for helping thinking: while I chat with AI, it generates a mind map of the conversation, so I can see the threads of my own thought. This direction has since been merged into my company’s project Blueprint, as one very small component — mind-map generation. It can’t be a long-term personal project, and the other goal — making revenue — would be hard there too.

VPN Connector. A toy I built for myself: the VPN tool I’d bought made connecting awkward, so I wrote my own VPN utility. But it can’t become a project — just a little tool for my own use.

Reddit Insight. This one ran for quite a while: mining demand from various subreddits, classifying and collecting it, split into B2B and B2C. It went on for some time but I didn’t keep it up. What I found: Reddit is flooded with people posting fluff — the people genuinely posting needs are far fewer, and huge numbers of people use Reddit to market and promote their own projects, with endless repetition. Extracting real demand means filtering out enormous noise. The other problems were, as always, conversion-to-payment and target users.

E-book converter. Converts PDF and EPUB into MOBI. I use it all the time, and it’s already live — on my book-translation site.

Video editing tool. Last year I wanted to build a video editing tool: script generation, then AI-driven cutting and assembly. I even asked our intern at the time to help me build it, but it never stuck.

Trend prediction (Trading). Even earlier, Chunjun and I built a project for trend prediction — a model predicting the rise and fall of large-cap stocks, pulling data from AKShare and using regression models, SVM, and the like. But I found the core of this is data: realistically you have to pay for good data, and I never made that investment. It stayed an experiment — I lacked the resolve.

What actually moved me

After all these attempts, looking back now: the only project I truly finished and delivered is the AI book-translation site. It’s live on the public web, it can be found on Google, it gets exposure every day — not much, one or two hundred impressions a day — and people do click. That touched me: it is real feedback, proof that demand exists. And that is what pushed me to reflect.

Three lessons

First: direction. The direction has to be one I know well: something I’m deeply familiar with, willing to persist with, and actually use in daily life. Then the product gives me strong positive feedback on its own, I have a visceral sense of whether it’s good or bad, and it’s a product I would use myself. Not like WYRD — something that was never mine to like or use.

Second: how to build. This is the more important realization from the past two days: don’t rush into building the product. Find users first. Do the research. Find the first batch of target users — know where they are, how to contact them, how to reach them — then organize what you learn, and know exactly who you will send the first release to once it’s ready.

Third: marketing first. Prepare the marketing materials from day one — marketing first, product second. Get Discord, Twitter, Reddit, Xiaohongshu, Zhihu, and a WeChat official account set up in advance. A WeChat official account isn’t a great marketing channel; but Xiaohongshu, Zhihu, Twitter, Discord, and Reddit are. Plus non-English community forums, and GEO: doing SEO inside ChatGPT, or advertising inside ChatGPT.

Persisting: keep a log

And then there’s persistence — continuing to do the work. How do you persist? One very good way is to record. Write down each day’s thinking, what you did, your lessons and reflections — as a log. Look back later: when you’ve already done this much, giving up would be a shame, wouldn’t it? A regret. That itself gives you more courage and more motivation to keep going. You need a sustained source of drive from both inside and outside: internally, the daily record that encourages you not to quit; externally, user feedback that gives you positive reinforcement and powers continuous iteration and continuous polish. That is how you keep going.

Where things stand, and today’s work

There’s already a foundation: my AI book-translation site has some traffic. I can build on it to drive more traffic and lay the groundwork for expanding into follow-up products.

Today I mainly did five things:

1. Checked the site’s data in Google Search Console: there are 404 page errors — unresolved for now; I’ll fix them tomorrow.
2. Optimized the Tools page: previously, converting PDF books to MOBI or EPUB produced terrible results — it went straight through OCR with no LLM post-processing, so the final books read poorly. This round optimized that part, with the GLM model as a fallback; it also now supports larger PDF files and emails the results to the user. The issue I see: LLM token consumption invites freeloading, so rate limiting may be needed — let each user process one file per day. This part of the site exists mainly to drive traffic, so I’ll try it free for now: one file a day, and beyond that they pay. Optimization isn’t fully done; once it is, the page itself needs adjusting — how to make this page stand out, so searches can find it like those free tools do.
3. Deep Research comparison: from yesterday into today, for the foreign-language book reader client I’ve been thinking about, I ran two Deep Research reports — one with Claude, one with Gemini. Overall, Gemini’s Deep Research did better than Claude’s.
4. Competitor research: based on the Google results and the deep research, I downloaded all kinds of book-reading client apps and found the product experience genuinely poor — most are really bad, and almost none are built with a proper experience for the specific need of reading foreign-language books. The closest one is Yingyue Reader (英阅阅读器), built by an indie developer, dedicated to reading English-language books — fairly close to my ideal product shape, though its experience still falls short of WeChat Reading.
5. Reddit promotion attempt: I meant to create a subreddit for promotion and marketing, but I’m not familiar with Reddit’s rules, so it didn’t succeed.

What’s next

The coming stretch is mainly product research and user research: going through every social platform, every app platform, every place, to understand the demand and the users — where the users are, who they are, and how I will eventually reach them. I need to find the first batch of target users. At a 1% conversion rate, if I want 100 people, I first need to find 10,000 — people I can message, people I can reach. My goal is to find those 10,000 first.