/notes/ ·
Three and a half months on one problem
I am a product-first person. I have always believed that if the product is the best one in the market, it wins without much of a sales pitch. That belief only works if you are willing to go and solve the part everyone else routes around.
In the spring I locked myself in and worked on nothing but Selda's hardest problem: write a message that sounds like the person sending it, in their voice, and have it convert.
Three and a half months went to that and nothing else. Vector databases, prompting, prompt tuning, every combination I could think of. Most days I opened the output and it was worse than the day before. Em dashes everywhere. That flat, agreeable tone you can smell in the first line, the one that makes a reader delete a message before they finish it. It was fairly miserable, honestly.
What eventually moved it was reading how people actually write, from every source I could find. Not marketing copy. Ordinary writing. There is depth in it, and the structure is unclear in a way a machine keeps wanting to tidy up. The mess is the signal. Once I stopped removing it, the messages started sounding like someone.
Now they sound like a person, in your voice, and they convert. I know because I have sold with them myself, over email. And they are not prompted into existence anywhere you could copy.
Being second into a market decides nothing. Take the hardest problem, solve it properly, then take the next one. A product built that way does not need much manual selling.