FateTell, a team of 5 people, uses Bazi (Eight Characters) to tell fortunes for overseas users. A single destiny report is priced at $39.9. According to the founder's public sharing, the repurchase rate is 38.7%—those in the consumer goods industry know what this number means; many brands struggle to even reach 20%. Seventy percent of the income comes from subscriptions, making it a high-frequency, essential demand.
This is not an isolated case. Overseas, Starla launched a few months ago and grew through TikTok, with a peak daily income of $124,000 and total revenue exceeding $2.5 million. The Indian astrology platform Astrotalk is even more exaggerated, with revenue exceeding 1 billion RMB in fiscal year 2025, an 85% year-on-year increase. The "Life K-line Chart" project, which quantifies Bazi into stock K-lines, reached 3.23 million views for a single post, causing the server to crash.
To be honest, I initially thought this was just metaphysics. Later, I ran the Bazi for myself and those around me, and the accuracy in the general direction was sixty to seventy percent, which shocked me—I'll explain this in detail later.
The money is real. The reason this business works is simple: the cost for AI to generate a numerology report is close to zero, while anxiety is recurrent—people check when their career is rough, when relationship problems arise, or at the end of the year for next year's fortune. More importantly, no one screenshots their financial app's annual report to share on social media, but if AI generates a destiny chart for you, you're likely to screenshot and post it.
Zero supply cost, recurrent demand, and spontaneous user sharing—few businesses have all three variables. The global astrology app market is valued at $5.69 billion this year and is expected to reach $11.7 billion by 2030. The same product sells for 9.9 RMB domestically but $9.9 USD overseas, a seven-fold difference in gross margin.
But most AI fortune-telling products don't last three months.
Why They Don't Last Three Months
A good track doesn't mean you can make money easily.
If you send your Bazi to ChatGPT, what comes out? "You will have help from influential people in your career," "Financial luck will be suppressed before rising," "Marriage is better late than early." Correct nonsense—words that apply to everyone mean nothing. Users know it's inaccurate after one try and won't return, let alone have a 38.7% repurchase rate.
A single prompt only calls upon the AI's general knowledge. It knows about the Five Elements, but it doesn't know how to adjust the climate for Wu Earth born in the month of Shen according to the Qiong Tong Bao Jian, nor how to judge the auspiciousness of Liang Xiangrun's Sheng-Wang-Ku years, or how to arrange the monthly flow using Yuan Shushan's sixteen-character method.
Even more fatal are chart arrangement errors. In actual use, I've encountered:
- Si and Chou being judged as a partial Metal combination. Liang Xiangrun explicitly states that partial triple combinations must include Zi, Wu, Mao, or You (the four cardinal positions); Si and Chou without them do not count. One misjudgment changes the entire nature of the chart.
- Monthly flow ratings reversed by two levels. A month that should be marked "high risk" is marked "stable."
- Great Luck cycles being split. "The first five years look at the Heavenly Stem, the next five look at the Earthly Branch"—Ren Tieqiao explicitly opposed this in Di Tian Sui.
The commonality of these errors: AI hasn't read the original texts and relies on general knowledge to apply formulas, failing on details. Users might not notice the first time, but after two or three uses, they find contradictions and never return.
The 38.7% repurchase rate isn't because the marketing is good; it's because the product is actually somewhat accurate. My own experience was the same—once I realized it wasn't just talking nonsense, I couldn't help but keep using it.
What I Did
I wanted to find out one thing: if AI actually reads the original classic texts instead of relying on general knowledge, how accurate is Bazi?
So, I used Claude Code to read through over thirty numerology classics, built a 44MB knowledge graph, and finally made it into a one-click installable Claude Code Skill.
Digitizing the Classics
This wasn't just throwing a few e-books at the AI. Most of these classics only exist as scanned images; there are no ready-made electronic texts. Some are vertical traditional Chinese from the Republican era with low OCR accuracy, resulting in many typos and missing words. I had to proofread them word by word against the original scans.
The installation package distributes 17 public domain classics: 13 for Bazi, including Zi Ping Zhen Quan, Di Tian Sui Chan Wei, Qiong Tong Bao Jian, San Ming Tong Hui, and Xu Lewu's four commentaries; for Zi Wei, there is the Ming Daozang version of Zi Wei Dou Shu Quan Shu; plus two by Yuan Shushan and one on Liu Ren. The full list is in the README.
The complete works of Liang Xiangrun, three books by Wei Qianli, and works by contemporary masters like Wu Junmin are protected by copyright and not distributed—but the methodological notes I extracted after reading them are in the package. These notes serve as the operating manual for the AI when arranging charts.
I wrote notes after reading every book word by word. Numerology rules are scattered across dozens of works, and the same concept often contradicts itself between authors. If you don't read them all, you can't tell who is right and under what conditions. Nine core notes totaled over 150,000 words.
The Chart Arrangement Process
Given a birth time, the system follows a complete 15-step process.
First, the basic engineering: convert the Gregorian birth time to True Solar Time (different cities have longitude differences; without correction, it could be off by an hour, ruining the whole chart), arrange the Four Pillars, list the Hidden Stems, arrange the Life Palace, and judge the strength of the Day Master.
Then, the core analysis: check Qiong Tong Bao Jian to see what climate adjustment this Day Stem needs in this month (winter cold water needs fire to warm the chart, summer dry earth needs water to moisten it), determine the Structure (whether the Month Command's essence penetrates the stems determines the whole structure), and identify the Useful God (Yong Shen). These steps are the hardest part of numerology, where different schools diverge most.
So, from here, four schools vote in parallel. Not just any four—I tried more initially but found some methodologies overlapped heavily. I narrowed it down to the four with the greatest methodological differences and complementary dimensions:
The four schools run independently: four-way agreement = high confidence, three-way = medium confidence, two-two split = low confidence, marked for verification. The logic is the same as ensemble voting in machine learning—single model accuracy is limited, but voting significantly improves it.
Finally, a layer of cross-verification between Bazi and Zi Wei Dou Shu is added. Bazi looks at Heavenly Stems and Earthly Branches (time dimension), while Zi Wei looks at the twelve palaces and stars (spatial dimension). These two completely independent systems are compared across dimensions like wealth, career, marriage, and health—only when both point to the same conclusion is it considered credible.
Several iron rules govern the process: no guessing (say "unknown" if unsure), no reverse-fitting (don't force explanations based on known facts), no splitting Great Luck cycles, and partial triple combinations must include the four cardinal positions. Any violation renders the analysis void.
Running Run Run Shaw's Bazi
Methodology is useless without results. Let's look at what the system produced for Run Run Shaw (Shao Yifu), born Nov 19, 1907, Chen hour, Ningbo, lived to 107, the godfather of Chinese entertainment.
The system arranged the Four Pillars:
Ren Water born in the month of Hai, gaining command and rooted, strong Day Master. The Month Command is Bi Jian (Friend), so no structure is taken from it; Shi Shen (Eating God) penetrates the Hour Stem, Zheng Cai (Direct Wealth) penetrates the Year Stem → Eating God Producing Wealth Structure. Checking Qiong Tong Bao Jian, Ren Water in the month of Hai needs Ding Fire for climate adjustment, and the Year Stem happens to be Ding Fire. The basic analysis is undisputed; all four schools agree.
The real discovery starts with the Heavenly Stem chain. The four stems form a complete cycle:
Xin (Resource) → Ren (Self) → Jia (Eating God) → Ding (Direct Wealth)
Resource produces Self, Self produces Eating God, Eating God produces Wealth—the four rings are connected in one breath. In human terms: others give him resources (Resource produces Self) → he transforms them into works (Self produces Eating God) → works turn into money (Eating God produces Wealth). Making movies makes money → the money is invested to make more movies; the flywheel is naturally connected.
Deeper level: Ding-Ren combination transforms into Wood (Auspicious). The Year Stem Ding Fire and Day Stem Ren Water combine to transform into Wood—which in this chart is Eating/Hurting, the Auspicious God. Other people's charts often have Wealth combining with the Self to transform into an Inauspicious God, where money brings trouble. Run Run Shaw is the opposite: money comes → becomes creativity → creativity becomes more money. A virtuous cycle of wealth.
Earthly Branches have no clashes, punishments, or harms. There are no violent confrontations between Wei, Hai, Shen, and Chen; no major upheavals in his life—living to 107 is clearly written in the branches.
Looking at the Great Luck cycles compared to his life trajectory:
From age 34 to 73, 40 consecutive years of Wealth and Eating/Hurting luck, exactly corresponding to the peak period from establishing Shaw Brothers to TVB dominating Hong Kong.
All these analyses were first produced blindly by the system and then verified against historical facts. If verification fails, the framework is modified; we don't force explanations.
Run Run Shaw is a public figure, but I also did blind tests for myself, family, and friends—providing only birth time and location. The hit rate for the general direction is indeed around 60-70%. Some conclusions seemed unlikely at first, but upon checking with the individuals, they were accurate. As a 2002-born STEM student who studied data science, telling you Bazi is 70% accurate feels absurd, but repeated testing shows it is so.
How to Use It
Two ways, depending on your needs. If you just want to experience it, copy the prompt to any AI. For the full system, install the Claude Code Skill.
Method 1: Enhanced Prompt (Works for all AI)
Copy and send this to ChatGPT, Claude, or DeepSeek. It includes the four-school voting and verification mechanism:
You are a professional researcher proficient in traditional Chinese numerology. Your analysis must use four methods simultaneously and determine a consensus through a voting mechanism:
Four Methods:
1. Xu Lewu Method: Structure-centric, climate adjustment takes precedence when urgent. Ref: Zi Ping Zhen Quan, Qiong Tong Bao Jian
2. Liang Xiangrun Method: Five-filter sorting (Sheng-Wang-Ku years > Net of Heaven and Earth > Fu Yin/Fan Yin > Day Master offending Year Lord > Structure). Ref: Zi Ping Ji Chu Gai Yao
3. Yuan Shushan Method: Arrange Life Palace + Small Limit + 16-character method. Ref: Ming Li Tan Yuan
4. Wei Qianli Method: Eight-step judgment + Great Luck corner theory. Ref: Ba Zi Ti Yao, Qian Li Ming Gao
Iron Rules (Violation = Invalid Analysis):
- No guessing: Say "uncertain" if unsure.
- No reverse-fitting: Blind analysis first, then verify.
- No splitting Great Luck cycles: Stems and branches are inseparable.
- Partial triple combinations must include Zi/Wu/Mao/You.
- Each conclusion must cite its source.
Please analyze based on the info provided:
[Personal Info] Date, Time, Gender, Location...
[Process] 1. Arrange pillars (True Solar Time) 2. Identify Day Master 3. Month/Strength/Structure/Climate 4. Independent analysis by 4 schools 5. Voting consensus 6. Predict past events for verification 7. Future trends 8. Summary.
Method 2: Install Claude Code Skill (Full System)
A prompt makes AI "pretend" to know numerology. A Skill is different: 17 original classics and 150,000 words of notes are in the package; the AI actually looks up the original texts.
The package is 59MB (zip), containing 61 files, including the classics and:
9 core methodology notes. These are the operating manuals for the AI. It checks the notes to set the framework and then flips through the original texts for details.
Famous Cases. Run Run Shaw, Mao Zedong, etc., for verification.
44MB Knowledge Graph. 45,895 entities, 108,184 relations. It allows for cross-book retrieval—asking about a specific concept will pull relevant paragraphs from multiple classics in seconds.
Installation
Unzip and run ./setup.sh. Choose Level 1 (Lightweight, no dependencies) or Level 2 (Advanced, includes Knowledge Graph).
Once installed, just talk to Claude Code:
"Arrange Bazi for me: May 15, 1990, 10 AM, Male, Beijing"
It will follow the 15-step process, citing sources for every conclusion down to the specific paragraph.
Boundaries
Liang Xiangrun said it most directly: "Numerology has a probability of about 60 to 70 percent."
My tests confirm this. The same Bazi in different cities or families can result in very different paths. Bazi describes potential space and risk windows, not a script. Anyone claiming "100% accuracy" is a liar.
But 60-70% is enough. The Google Drive link is in the comments.
