How Custom Instructions Work

🔄 In Lesson 1, you saw that custom instructions make every AI conversation smarter. But how do they actually work under the hood? Understanding the mechanics helps you write better instructions — and troubleshoot when things go wrong.

The Message Hierarchy

Every AI conversation has layers. When you send a message, here’s what the AI actually sees:

[System Message]     ← Custom instructions live here (highest priority)
[Conversation History] ← Previous messages in this chat
[User Message]       ← What you just typed

Custom instructions sit at the system level — the very top. The AI reads them first, before anything else. This is why they’re so powerful: they set the ground rules for the entire conversation.

But here’s the thing most people miss: custom instructions aren’t magic overrides. They’re strong suggestions. If your instruction says “always use bullet points” but you ask for a cover letter, the AI has to reconcile both. Usually, the specific request wins over the general instruction.

✅ Quick Check: Your custom instructions say “Keep responses under 200 words.” You then ask the AI to “write a detailed 1,000-word analysis.” Which instruction wins? (Usually the specific request wins. The AI will attempt the 1,000-word analysis. This is why good custom instructions include exceptions: “Keep responses under 200 words unless I ask for longer output.”)

Platform Character Limits

Every platform gives you a different budget for instructions:

The takeaway: ChatGPT’s basic Custom Instructions are the most constrained at 1,500 characters per field. Claude and Gemini give you more room. But constraints aren’t bad — they force clarity.

ChatGPT’s Two-Field System

ChatGPT splits custom instructions into two fields:

Field 1: “What would you like ChatGPT to know about you?”

• Your role, profession, expertise level
• Context about your work or projects
• Preferences about how you communicate

Field 2: “How would you like ChatGPT to respond?”

• Output format preferences (length, structure, tone)
• Things to include or avoid
• Language and style preferences

This split is actually clever. Field 1 is context (who you are). Field 2 is behavior (how the AI should act). Keeping them separate helps the AI treat them differently.

Example:

• Field 1: “I’m a product manager at a B2B SaaS startup. We serve mid-market companies. I work mainly with engineering and design teams.”
• Field 2: “Be direct and concise. Use bullet points for action items. When I ask for feedback, be critical — I don’t need encouragement, I need to find problems. Default to American English.”

Claude’s System Prompt Approach

Claude handles it differently. Instead of two fields, you get one system prompt — but with much more flexibility.

In Claude Projects: You set “Project Instructions” that apply to every conversation within that project. You can also attach documents as context.

In the Claude API: You set a system parameter. No character limit (within the context window), which means you can write extremely detailed instructions.

Claude’s secret weapon: XML tags. Anthropic recommends structuring long instructions with XML:

<role>You are a senior Python developer...</role>
<rules>
  - Always include type hints
  - Use docstrings for all functions
  - Prefer list comprehensions over loops
</rules>
<output_format>
  Start with a brief explanation, then code block
</output_format>

XML tags help Claude parse complex instructions cleanly. They work on other platforms too, but Claude is specifically optimized for them.

✅ Quick Check: Why might you choose Claude Projects over ChatGPT Custom Instructions for a complex writing project? (Claude Projects let you attach documents as context and write longer, more detailed instructions. ChatGPT’s 1,500-character limit per field forces you to be brief, which is fine for general preferences but limiting for project-specific context.)

What Happens When Instructions Conflict

This is the question everyone hits eventually: what if your custom instructions say one thing and your prompt says another?

The hierarchy (in practice):

1. Explicit user request in the current message → strongest
2. Custom instructions / system prompt → strong
3. Default model behavior → weakest

So if you tell the AI “respond in French” in your instructions but type “answer in English” in your message, you’ll get English. The specific, recent request overrides the persistent instruction.

Best practice: Write instructions with built-in flexibility. Instead of rigid rules, use defaults with exceptions:

• “Use bullet points by default. Switch to prose when the task requires it.”
• “Keep responses concise unless I ask for detailed analysis.”
• “Respond in English unless I write in another language.”

Token Economics

Custom instructions consume tokens from your context window. On ChatGPT with 3,000 characters of instructions, that’s roughly 750 tokens — out of a context window of 128,000 (GPT-4o). That’s less than 1%.

But for short conversations, those 750 tokens are a bigger percentage of what the AI processes. The takeaway: keep instructions focused. Every sentence should earn its place.

The efficiency test: Read each line of your instructions. If the AI would behave the same without it, delete it. Instructions like “be helpful” or “be accurate” are wasted space — the AI already tries to do both.

Key Takeaways

• Custom instructions sit at the system level — they’re read before any of your messages
• ChatGPT uses two fields (context + behavior); Claude uses one flexible system prompt with XML tags
• Character limits vary: ChatGPT is most constrained (1,500/field), Claude and Gemini are more generous
• When instructions conflict with prompts, specific recent requests usually win over general instructions
• Write instructions with built-in flexibility (“by default” and “unless”) to handle edge cases
• Every instruction consumes tokens — cut anything the AI would already do without being told

Up Next

You understand the mechanics. In Lesson 3, you’ll write your first custom instructions using the RISEN framework — a structured approach that works across all platforms. You’ll have working instructions in 15 minutes.