DEV Community

Ian
Ian

Posted on

Building for a market where you can't read the reviews

I ship a product primarily for the Japanese market. I don't read Japanese fluently. This is a description of the workaround, and the specific ways it fails.

Why Japan

Not strategy, mostly arithmetic. The category I'm in — wedding speeches — has a specific structural property in Japan that it doesn't have in English-speaking markets: there's an established list of words you must avoid (imikotoba / 忌み言葉), and everyone who gives a wedding speech is expected to know it.

That's a compliance layer. Compliance layers are where software is actually useful, because the value is in not making a mistake rather than in being creative.

In English, a best man speech has no equivalent list. The problem is "I don't know what to say." In Japanese it's "I don't know what I'm not allowed to say." Different problem, much better fit for a tool.

The workaround stack

1. Write in the target language via the model, then verify against rules — not by feel.

I can't tell whether a Japanese paragraph sounds natural. I can tell whether it contains 切る. So the architecture leans on checkable properties: character count, taboo-word absence, required structural elements (opening, self-intro, story, toast).

2. Publish the rules as data.

If the etiquette lives in prose I can't evaluate, it's unusable to me. As JSON with a test file, it's something I can assert against. That's the open dataset — born from my own inability to check my work.

3. Compete on measurements I can take.

I can't judge whether my Japanese reads better than a competitor's. I can measure that my page has 2,100 characters and theirs has 8,000. So content depth became my proxy metric, and it turns out to matter more than I expected.

4. Treat native-speaker review as a blocker, and be honest when it hasn't happened.

This is the uncomfortable part: the project's own spec says Japanese content needs native review before it ships. Some of it hasn't had that. I've mitigated with mechanical checks and with sourcing claims to citable pages, but mechanical checks don't catch "this is grammatically fine and culturally odd."

I'd rather name that risk than bury it, because a solo builder pretending to have solved a language they don't speak is how products get embarrassing in a market they can't see.

What specifically goes wrong

Numbers. A 3-minute Japanese speech is ~600–900 characters. A 3-minute English speech is ~350–600 words. If you write one rule for both, one of your locales is badly wrong and you won't notice, because everything renders fine.

Tone. English wedding speeches tolerate self-deprecating humour and mild roasting. Japanese ones are more formal; a joke that lands in Manchester reads as careless in Tokyo. The model needs different instructions per locale, not a translated version of the same instruction.

Trust signals. Japanese buyers look for different things. A US buyer reads testimonials. A Japanese buyer notices whether the operator is identifiable, whether there's a clear refund policy, whether the site looks maintained. I have no testimonials — I've had zero sales — so the refund policy and the visible maintenance have to carry it.

Payment expectations. Prices in Japan are quoted tax-inclusive. If your checkout adds 10% at the end, it reads as a bait price. This one cost me a real fix.

The uncomfortable question

Should I be building for a market I can't read?

The honest answer is: it depends on whether the product's value is checkable. Mine is — it's structure, length, and word avoidance. If it were "write something beautiful in Japanese," I'd be a fraud with an API key.

The boundary I try to hold: I can build the machine. I can't be the taste. Where taste is required, the product needs a human who has it, and until then the right move is to say so.

If you've built for a market you don't speak the language of, I'm curious whether you found the same line, or a different one.

(The product: Speech Kobo. The dataset that came out of this constraint: GitHub.)

Top comments (0)