🎯 Core Goals
- Show the same dual-view as 3.2, but now the bundle always starts with hidden layers.
- Understand the three prompt layers (Provider / App / Session).
- Demystify Custom GPTs, Claude Projects, and Gemini Gems as pre-packaged sandwiches.
Before you type your first message, the LLM has already received hidden instructions. Every bundle has invisible layers on top — and some of them you can’t touch.
👁️ Visuals & Interactives
The Full Sandwich — Hidden Layers
Before you type a word, the bundle already has invisible layers on top
Press Send to start...
↑ These two layers are always here, before your first message.
📝 Key Concepts
Three layers in every bundle:
- Layer 1 — Provider Prompt (You can’t see or change this): Set by the AI company (Anthropic, OpenAI, Google). Contains safety rules always active: “Don’t produce harmful content,” “Never pretend to be human when sincerely asked.”
- Layer 2 — App / Custom Pre-prompt (Set by the app builder): Defines the product’s personality, tone, and knowledge. “You are TechCorp support.” Makes a chatbot feel specialized.
- Layer 3 — Your Session Context (Yours to set): Repeated preferences you inject per session: “Always respond in bullet points,” “I’m working in Python 3.12.”
The Hidden Cost of Context: Now that you know sending a message often means sending a “bundled” message, here’s the catch: since LLMs don’t “remember” previous turns, you re-send the entire history (the bundle) with every new reply. This means the 10th message in a chat costs significantly more to process than the 1st one!
📦 Custom Spaces — Pre-packaged Sandwiches
Features like Custom GPTs, Claude Projects, and Gemini Gems bundle three things:
- System Prompt: Specialized instructions (personality).
- Knowledge Docs: Reference files (company policies, product specs).
- Tools: Specific capabilities (web search, code interpreter).
Why use them? Set the instructions once, then just use the chat. No need to re-explain context every session.
Business use: Companies build internal Projects pre-loaded with brand voice, legal policies, and private data — so every employee starts with the right context.
Why does Claude feel different from ChatGPT? Their Provider Prompts (Layer 1) and default pre-prompts are completely different. Same concept, different instructions.
What allows a Custom GPT or Claude Project to feel specialized without you explaining anything?