Key Moments
Stanford CS193T Thinking with AI | Autumn 2026 | Core Prompting
Stanford OnlineKey Moments
Prompting AI is like training a new hire: be clear, specific, and avoid ambiguity to get useful results. Otherwise, you risk relying on the AI's "hallucinations" or losing your own critical skills.
Key Insights
Treating AI as a new hire requires clear goals, specific tasks, and acceptance criteria, just like managing a human employee.
The 'interview me' pattern is a method to elicit necessary questions from AI, ensuring all requirements are met before task execution.
Prompt postmortems can be used to create reusable, detailed prompts that capture context and ensure consistent, high-quality results.
Over-reliance on AI can lead to a degradation of human skills, akin to losing the ability to navigate without GPS.
System prompts, like those used by Anthropic, are hidden instructions that guide AI behavior and demonstrate transparency when made public.
Context window management is crucial; starting a new chat for a new topic prevents token waste and potential 'context rot'.
Treating AI like a new hire requires clear direction
The core principle of effective AI prompting is to treat the AI as a capable, yet new, employee. This means providing clear, detailed instructions, much like a scope of work document for a human. Just as a manager sets expectations for a new hire, users must articulate their desired outcomes precisely. This involves defining who the writer is, who the target audience is, the purpose ('why') of the task, and any constraints ('how'). A good analogy is that of a project manager who clearly defines goals, breaks down tasks, sets acceptance criteria, and provides intuitive guidance. Without this clarity, the AI, like a new employee, may produce results that don't meet expectations, leading to frustration. The key takeaway is that the AI adapts to your instructions, not the other way around, and requires human oversight to achieve optimal results.
The 'interview me' pattern elicits crucial information
When you're unsure of all the necessary details for a task, the 'interview me' pattern is a powerful technique. Instead of providing a vague request, you instruct the AI to ask you a series of clarifying questions (e.g., 'Ask me at least 10 questions to understand my needs'). This prompts the AI to surface potential gaps in your initial request. For example, if planning an event, the AI might ask about the organization type, event scale, budget, and desired outcomes. This collaborative questioning ensures that all requirements are considered before the AI begins its work, significantly reducing the chance of errors or misinterpretations. It transforms the AI from a passive executor into an active partner in requirement gathering.
Prompt postmortems create reusable, detailed instructions
A valuable technique for improving future interactions with AI is the 'prompt postmortem.' After successfully completing a task, especially one that required multiple iterations, you can ask the AI to generate a detailed, standalone 'prompt' that replicates the successful interaction. This prompt essentially captures the context, successful exchanges, and final instructions, allowing you to achieve the same result from the first attempt in the future. This is particularly useful for recurring tasks, such as generating reports in a specific format or coding specific functions. By saving these detailed prompts, you build a library of effective instructions, saving time, computational resources (tokens), and ensuring consistency. This practice also has environmental benefits by reducing the number of tokens processed.
Trust and observability are key to catching AI errors
The lecture emphasizes the importance of trust and observability when interacting with AI, particularly concerning potential 'hallucinations' or confident mistakes. AI models do not 'know' in the human sense; they predict the next token based on their training data. When asked to reflect on their own processes or confidence, they generate text that sounds plausible but lacks genuine self-awareness or access to the underlying computational states. Therefore, instead of asking 'How confident are you?', it's more effective to ask for evidence and sources. Users should critically examine provided sources, as AI can generate plausible-sounding but inaccurate or fabricated information. Observability, the ability to understand the system's operation, can be enhanced by explicitly asking the AI to document its steps or summarize its reasoning process within the chat, making the AI's work more transparent and verifiable.
AI interaction can lead to skill degradation if not managed
A significant concern highlighted is the potential for human skill degradation through over-reliance on AI. The analogy of losing the ability to navigate without GPS illustrates how easily critical skills can atrophy. Experts interviewed noted that they sometimes realized they had forgotten how to perform routine tasks, like coding, because the AI had been doing it for them. To combat this, the lecture advises actively maintaining and protecting essential skills. This involves not just using AI as a tool but also ensuring you understand *how* it performs tasks and retaining the ability to do them independently. Keeping a record of AI-generated solutions with explanations on how to replicate them can help preserve knowledge and prevent the subtle, gradual loss of expertise, which can go unnoticed without a specific trigger event.
Context window management is crucial for efficient AI use
Effectively managing the AI's context window—the amount of text it can process at once—is vital. The lecture uses the metaphor of 'new wine in new wineskins' to emphasize that new topics warrant new conversations. Mixing disparate subjects (e.g., relationship advice, curry recipes, calculus) in a single chat wastes tokens and dilutes focus. While AI models often prioritize recent prompts, irrelevant historical context can still lead to inefficiencies and potential 'context rot,' where the AI's responses become biased by old, unrelated information. Starting a new chat for a new topic is the most efficient way to manage resources and maintain focus, ensuring the AI's attention is solely on the current task.
Resource management and appropriate tool selection are key
Using AI effectively also involves smart resource management. This means selecting the right model, effort level, and settings for the task at hand. A simple task, like proofreading a short text, doesn't require the most powerful and computationally expensive model. Similarly, avoid overly cumbersome input methods; pasting text is more efficient than using image recognition to extract text. Crucially, AI is not always the best tool. For simple calculations, use a calculator; for finding specific text, use 'Control + F.' Overusing large language models for tasks that simpler tools can handle increases computational cost and energy consumption unnecessarily. Choosing the right tool for the job is fundamental to efficient and responsible AI use.
Maintaining your independent voice and critical thinking
The lecture concludes by stressing the importance of maintaining one's independent voice and critical thinking. The AI, left unchecked, can follow the 'path of least resistance,' essentially reflecting the biases and averaged opinions from its training data ('the illusions of others,' as Audre Lorde put it). This can lead to a loss of self and individuality. Therefore, it's imperative to assert your own perspective, desires, and will. This is not just a philosophical point but a practical necessity to ensure the AI serves your goals rather than imposing its own learned biases. By actively engaging, questioning, and documenting AI interactions, users can ensure they remain in control, learn from the process, and preserve their own cognitive abilities.
Mentioned in This Episode
●Software & Apps
●Companies
●Books
●People Referenced
Core Prompting Strategies for Effective AI Interaction
Practical takeaways from this episode
Do This
Avoid This
Common Questions
Interactive chat AI involves moment-to-moment back-and-forth, similar to a conversation. AI agents, on the other hand, can be given a task and then work autonomously for extended periods, allowing you to focus on other work.
Topics
Mentioned in this video
Mentioned in the context of the 'lost in the middle' phenomenon within a long context window, likening it to being lost in the middle of life's journey.
Used as an analogy to illustrate the importance of having a clear goal when interacting with AI. The Cheshire Cat's advice to Alice highlights that without a destination, any path will do, emphasizing the need for specific objectives when prompting AI.
An AI model developed by Anthropic. Its system prompt is used as an example of direct and precise language. The speaker also mentions using Claude for coding tasks and debugging.
The platform used for the course assignments and exercises, allowing users to choose any AI provider's model for their work.
An AI model mentioned alongside OpenAI and Claude, particularly in the context of its 'thinking' or 'chain-of-thought' feature. The speaker demonstrates how Gemini's internal thought process is presented differently compared to Claude.
Mentioned in the context of AI models and their different settings, specifically in relation to the 'thinking' or 'chain-of-thought' feature, alongside Gemini and Anthropic.
A company whose CEO, Jensen Huang, is cited for a statistic that over half of their chips are used for inference rather than training, highlighting the significant computational resources involved in AI.
The company that developed the Claude AI model. They are mentioned for their transparency in making their system prompts public, which is presented as a good example of responsible AI practices.
More from Stanford Online
View all 172 summaries
78 minLLMs for Agent Builders: Context Engineering | Stanford CS329Z Lecture 2 with Michael Ryan
73 minStanford CS193T Thinking with AI | Autumn 2026 | Introduction to AI
70 minStanford CS329Z Introduction — What Are Agentic Systems?
65 minStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | Extra Session - Particle Filter SLAM
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free