Conventional wisdom for a small AI product: keep your prompts secret. They're the moat.
I went the other way and published the dataset behind my product's checks as CC-BY-4.0. Here's the reasoning, and where it might be wrong.
What the product actually is
It generates wedding speeches. Nine questions in, three drafts out. The interesting part isn't the LLM call — that's an API call and a JSON schema. The interesting part is that Japanese wedding speeches have a list of words you must avoid, called imikotoba (忌み言葉), because they suggest separation or endings.
Without that list, the model writes fluent text that would make someone's grandmother unhappy.
Why I published the list
1. It isn't actually a moat.
The list is cultural knowledge. It exists in books, in wedding planning sites, in every Japanese etiquette article. I didn't invent it; I compiled it. Treating compiled public knowledge as a secret is a category error — and worse, it means the data quality never improves because nobody can check it.
2. Hiding it made my own code worse.
When the data lives inside a prompt string, you can't test it. You can't diff it. You can't notice that 終わる/終わり needs splitting on the slash or the matcher silently misses every variant.
The moment I pulled it out into a typed JSON structure with a test file, that bug surfaced in five minutes. It had been latent for weeks.
3. It's a signal, not a giveaway.
Someone who wants a wedding speech is not going to build a matcher, write the prompt, build the UI, and wire up payments — because they found a 4KB JSON file. The overlap between "will use the dataset" and "would have paid $19.90" is approximately zero, and the first group is developers who will mention the source.
What it cost
Honestly? Almost nothing measurable. The repository has a modest number of stars. Nobody has cloned the product.
What it did produce: a reason to write three technical posts, a plausible claim to competence, and the kind of bug-finding that only comes from treating data as data.
Where this advice would be wrong
If your moat is proprietary data you actually collected — user behaviour, hand-labelled corpora that took months, something competitors cannot reproduce — publish that and you're giving away the business.
The test I'd apply: could a determined competitor reproduce this in a weekend? If yes, it isn't a moat and hiding it just makes your own engineering worse. If no, keep it.
Imikotoba is reproducible in a weekend by anyone who reads Japanese. So it was never a moat. I just hadn't said that out loud.
The concrete version
- Dataset: github.com/johnsiebyron-stack/imikotoba
- The prompt itself: still private, because it encodes structure and tone decisions that are product judgement
- The product: speechkobo.com
Line I'm drawing: facts are open, judgement is private. The list of taboo words is a fact. Which one to use instead, in what order, at what length, for which speaker — that's the product.
Still not sure it's right. If you've shipped something with a "moat" that was really just compiled public knowledge, I'd like to know how it went.
Top comments (0)