Stanford CS193T Thinking with AI | Autumn 2026 | Managing Sources
Stanford Online
Saving preferences 0:05
You can store preferences in the major AI tools so you do not have to repeat them in every prompt. In ChatGPT, the settings menu has personalization controls for voice, structure, emoji, and speed, plus a free text box that acts like a custom system prompt. Claude has a similar instructions box under settings and profile, where you can ask for less sycophancy or even add playful habits such as every third reply in pirate speech.
Memory and privacy 7:02
These tools also learn from past chats and can reuse facts you have disclosed, such as where you live, what work you do, or your age. ChatGPT lets you turn memory on or off and review what it has stored. Gemini can help you import memory from ChatGPT by giving you a prompt that asks ChatGPT to list its memories in a strict format, which is an example of a carefully specified scope of work prompt.
RAG Needs Care 21:01
RAG is not a license to dump anything into the context window. The quality of what you add matters because the system may lean harder on your stored sources than on general training data. Bad information can therefore sway the result more than you expect. The model does not always trust the RAG store over its training, though. Sometimes it accepts a bad source too readily, and sometimes it clings to training data when you want it to follow the document you provided.
Skills and Withholding 28:30
The same logic applies to saved skills and prompts. Claude skills and Gemini gems let you store detailed instructions for later use, and people share them online for tasks ranging from Stanford branding to game-like chat adventures. These tools sit in the context window rather than changing model weights. Strategic withholding can also help. If you want a sharper critique, frame a resume or essay as someone else’s work, use a fresh chat, and keep unnecessary facts out when privacy, security, or clean reasoning matters.
Brainstorming and token cost 42:01
When you ask an LLM for more ideas, do not show it your first draft if you want diversity. A single example can anchor the rest of the brainstorm. The talk also notes that token budgets usually count output tokens, including chain-of-thought reasoning, so long reasoning can still drive up cost even when the visible answer is short.
Prompt injection risks 45:01
The system prompt is not in a separate secure zone. It sits in the same context window as your own text, so an LLM cannot always tell them apart. That makes prompt injection possible when web documents are searched and pasted in. A page can hide instructions that tell the model to ignore earlier rules or take harmful actions, which is why web search, links, and email connectors need care.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

