← sources

Has anyone else noticed this? Claude writes prompts in a way that somehow leads to more prompts instead of actually achieving the goal.

source: reddit (reddit:ClaudeAI)published 2026-09-19 16:22ingested 2026-09-20 02:36

retained text

You give it clear inputs and ask it to create a prompt for an agent to do something, ask Claude to write a prompt for a specific task, and instead of just translating your requirements into a final prompt, it adds so much orchestration and meta-instructions that the resulting prompt starts steering the development in a completely different direction.  
    
You think you got a prompt for your task.  
You actually got a prompt that tells another agent how to think about your task, what process to follow, what decisions to make, and sometimes even what direction to take.  
  

So be careful when using Claude to write prompts.

  Read the generated prompt carefully, it can subtly change the actual objective.

document fetched 2026-09-20 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wkqgqo/has_anyone_else_noticed_this_claude_writes