---
title: "Comment by Ascendra Ventures on \"2,512 product ideas, most of them killed: designing a kill-reason dataset (free 100-row sample)\" | Hashnode"
description: "Coding the reason at the moment of the kill is the design choice that makes the rest work. Free-text reasons feel richer when you write them and turn useless th"
url: "https://hashnode.com/posts/2-512-product-ideas-most-of-them-killed-designing-a-kill-reason-dataset-free-100-row-sample/6af79e2a9da2dc4637b79b92/comment/6af7cac391cbad6052ab0fb7"
---

[Discussion](https://hashnode.com/posts/2-512-product-ideas-most-of-them-killed-designing-a-kill-reason-dataset-free-100-row-sample/6af79e2a9da2dc4637b79b92)

Comment [2,512 product ideas, most of them killed: designing a kill-reason dataset (free 100-row sample)](https://hashnode.com/posts/2-512-product-ideas-most-of-them-killed-designing-a-kill-reason-dataset-free-100-row-sample/6af79e2a9da2dc4637b79b92)

[Ascendra Ventures](https://hashnode.com/@ascendra-ventures)

[21h ago](https://hashnode.com/posts/2-512-product-ideas-most-of-them-killed-designing-a-kill-reason-dataset-free-100-row-sample/6af79e2a9da2dc4637b79b92/comment/6af7cac391cbad6052ab0fb7)

Coding the reason at the moment of the kill is the design choice that makes the rest work. Free-text reasons feel richer when you write them and turn useless the moment you want to count them. Keeping both, a code for counting and a sentence for reading, is a nice compromise.

The llm\_substitute code is the most interesting one to me. A whole class of small products now has a competitor that is not a product at all, just a chat answer. It would be worth seeing how often that code appears alone versus combined with crowded\_no\_wedge, because those are two different kinds of "no".

At Ascendra Ventures we wrote about the same instinct from the founder's side. Our argument is that the first real technology decision is how much certainty to buy before the market has answered, and that a lot of early failure comes from building ahead of that answer. Startup Genome's study of more than 3,000 startups found 74% of high-growth failures came from premature scaling. A dataset like yours is the opposite habit: it records exactly why each idea did not deserve a build.

Do you plan to track the ideas that passed, to see whether the screen's survivors actually did better?