Prompt Stacking
Published Aug 15, 2025
Prompt stacking is an informal term for the (invisible) compilation that happens between the time you submit your prompt and when it gets tokenized. Below is a description of what happens after you hit send and how you can tailor the LLM. Note that different platforms/tiers have different controls.
System: All LLMs have system level instructions that go with every prompt you send. This includes both customizable defaults and fixed guardrails.
Core: In ChatGPT (free and paid) this is under user > customize. For Claude paid, look under profile > preferences. These instructions set global rules, tone & output preferences for the LLM.
Project (ChatGPT Paid) / Workspace (Claude Paid): Project parameters, tone, output style, reference material in files.
Persistent memory (ChatGPT Paid & Claude Paid.): You can toggle this on/off. While on, it saves a memory when it deems it useful context long term.
Your prompt: The question.
Example: You can set a professional tone in the core and override it with a chatty creative type inside a fiction writing project.
Block | What it is | Example Input |
---|---|---|
System | Tone/output defaults and guardrails/safety set by developer. Guardrails outside user control. | "You are ChatGPT, trained to refuse disallowed content. Training date: April 2025. Guardrails: [safety rules]." |
Core | User-level personalization that applies to all chats. Overrides out-of-box style defaults set by developer | Prioritize well-respected sources over fringe. No sycophancy. |
Project/Workspace Instructions | Will enhance or override core. | Craft Soda Social Media: You are a creative collaborator in my online presence for Craft Soda Inc. We work together on... |
Persistent Memories/Memory | Personal context that persists across all projects | Content creator who works from home with three children under 10. |
User Prompt | The current message typed by the user, including constraints/facts. | Suggest 5 cute soda names for a blend of ginger, cherry and lemon. |
It all gets built and tokenized
|