System Prompts Reading & Discussion

​Frustrated with certain AI's behaviors (e.g., Claude Code's over-eagerness to fix what wasn't asked, GPT-5 getting stuck with em dashes), I decided to look into their system prompts to see what caused these behaviors.

​Questions I hope to discuss:

• ​Different system prompts for the same task (e.g., Claude Code vs. Cursor for coding). What do these differences mean for their behaviors?

• ​Differences in system prompts between the general and specialized versions (e.g., Claude vs. Claude Code)

• ​Guardrail approaches (e.g., between OpenAI and Anthropic)

• ​What can we learn from these system prompts to write better prompts? To work better with these tools?

​I'm very excited about agents, so I'll focus more on tool curation and instructions.

• ​This is NOT a jailbreaking guide!!!

• ​We won't be discussing anything confidential.

• ​Since system prompts change all the time, we'll focus on general principles.

​Q: Have these system prompts been confirmed by model providers?

​A: No. Anthropic released partial system prompts for Claude. I said partial because:

• ​They only show the general guidelines. The exact tool use (e.g. search_instructions) and safety instructions are missing.

• ​No system prompt for Claude Code.

​Q: Then how can you be sure that you have the correct versions?

​A: I'm not sure. However, I've done some cross-checking:

• ​With other published versions (from x1xhlol and elder-plinius)

  • ​Many prompts have changed from the previously published versions.

• ​With official system prompts.

• ​Reproduced across multiple attempts.

​Would love to discuss these with you!