AI and the Hidden Instructions Every AI Has

Behind every chatbot is a hidden system prompt — instructions the company gave the AI before you ever typed. That's why ChatGPT 'feels' different from Claude.

Some examples

• A system prompt sets tone, refusals, and persona.
• You can't see the system prompt directly in most chatbots.
• Custom GPTs let you write your own system prompt.
• The system prompt is why an AI may refuse certain topics.

Try it!

Open ChatGPT or Claude. Ask 'what are your default instructions?' Notice how vague the answer is — that's the hidden system prompt at work.

Every AI chat has two layers: the system prompt that sets the rules (you don't see it) and the user prompt that's what you type. Knowing this explains why ChatGPT and Claude act so differently with the same question.

How to use it

• Ask AI to explain what a system prompt usually contains
• Ask AI to describe how its own behavior is shaped by one
• Ask AI to compare 2 chatbots and guess their system prompt vibes
• Build your own system prompt for a custom GPT

Try it

Build your own custom GPT (or Project) with a 5-line system prompt. Compare its behavior to the default.

ChatGPT's 'You are a helpful assistant' is a system prompt. Claude's is much longer. Knowing they exist is how you start customizing AI for real.

Some examples

• ChatGPT 'Custom Instructions' replaces parts of the system prompt
• Claude Projects let you set a custom system prompt per project
• Poe lets you build a bot with any system prompt you want
• System prompts are why GPT in Cursor 'feels different' from ChatGPT

Try it!

Open ChatGPT settings → Personalization → Custom Instructions. Write 3 lines about how you want it to talk to you. Test the difference.

Every conversation with ChatGPT, Claude, or Perplexity starts with a hidden 'system prompt' — instructions the company wrote telling the AI how to behave. That's why the same underlying model behaves like 'helpful tutor' in one app and 'sassy friend' in another. When you build with the API, you write your own system prompt.

Some examples

• ChatGPT's system prompt is leaked online: it includes 'be helpful,' 'follow safety guidelines,' 'use markdown for formatting.'
• Custom GPTs and Claude Projects let you write your own system prompt — that's literally how they're built.
• Snap's My AI uses GPT-4 with a heavily customized system prompt that gives it the 'fun friend' personality.
• If you can write a clear 1-paragraph system prompt for your own GPT, you've done the same thing as 80% of 'AI app' startups.

Try it!

On chatgpt.com, click 'Explore GPTs' → 'Create.' The 'Instructions' box is the system prompt — write a 5-line spec for a useful tool ('always answer in 3 bullets,' 'never use jargon'). You just shipped a custom AI app.

Every AI you talk to has a hidden 'system prompt' — a long instruction the company wrote that tells the model how to behave (e.g., 'You are ChatGPT, a helpful assistant. Refuse to do X. Always cite Y.'). When you make a Custom GPT or Claude Project, you're writing your own system prompt on top. This is why the same underlying model (GPT-4o) can act totally different depending on which app you're using. Knowing this lets you build your own AI tools — and lets you understand why some chatbots feel friendly and others feel stiff.

Some examples

• Anthropic publishes Claude's full system prompt at anthropic.com/news — you can read exactly what behavior is hard-coded.
• OpenAI's Custom GPT builder is just a UI for writing a system prompt + uploading files — that's the entire 'magic.'
• When someone shares a 'jailbreak' prompt that seems to free the AI, they're trying to override the system prompt — the company logs and patches these.
• You can build a Custom GPT in 5 minutes (free in ChatGPT) with a system prompt like 'You are an AP US History tutor for a teen. Answer in Socratic style.'

Try it!

In ChatGPT, open the Explore GPTs page and pick any popular Custom GPT. Click 'View' to see the system prompt the creator wrote. That's the whole secret.