Andrew Ng's Prompting Course: The Key Gap Between AI Beginners and Power Users

Compared to when ChatGPT first launched in 2022, the way we interact with AI models has changed dramatically. When ChatGPT debuted in November 2022, users primarily engaged through single-turn Q&A — the experience felt more like a smarter search engine. By 2024–2025, the interaction paradigm has evolved into a combination of multimodal input (text, images, files, voice), multi-turn deep conversations, web search, code execution, and more. OpenAI's o1/o3 series introduced Chain-of-Thought reasoning, where models engage in extended step-by-step internal reasoning before producing an answer — this is the technical foundation behind what we'll discuss later as "giving AI enough time to think." This paradigm shift means users are no longer just "question askers" — they're "task orchestrators."

In his latest course, AI Prompting for Everyone, Andrew Ng states: Effective use of AI is one of the most impactful skills you can develop right now.

Yet many people who haven't reached the frontier of AI usage often feel frustrated when AI produces disappointing outputs. The problem usually isn't model capability — it's how the user prompts it. Drawing on Andrew Ng's core insights, this article breaks down the key gaps between AI beginners and power users.

Beginners vs. Power Users: Two Completely Different Experiences

Andrew Ng uses two typical profiles to contrast different approaches to AI. Many beginners treat AI like a search engine, asking only simple questions.

For instance, you might ask: "Does Taco Bell still have the Double Decker Taco?" — you'll get a simple answer, which is fine. But if you have more complex needs — say, you're buying a car — you can upload an entire set of documents to mainstream AI services like ChatGPT, Gemini, or Claude: vehicle specs, price quotes, insurance options, and then ask it to read everything carefully and think before responding with a thorough analysis.

This ability to "upload documents for AI to read" relies on the ever-expanding context window of large language models. Early GPT-3.5 supported only around 4K tokens (about 3,000 English words), while today Gemini 1.5 Pro supports 1 million tokens and Claude 3.5 supports 200K tokens. A larger context window means the model can "see" more material at once. Additionally, Retrieval-Augmented Generation (RAG) allows systems to first retrieve relevant document snippets from an external knowledge base and inject them into the prompt — effectively overcoming the physical limits of the context window and improving the accuracy and relevance of responses.

In these cases, AI might spend seconds or even minutes in deep thought before delivering a comprehensive comparison report. Ng admits this kind of usage has saved him enormous amounts of time. The core difference: beginners seek simple answers, while power users hand AI complex tasks and give it enough time to think.

Providing Rich Context: Think of AI as a Brilliant New Hire

AI power users learn to give AI the right context and background information — laying the foundation for high-quality responses. Beginners, by contrast, tend to use very short prompts and hope AI fills in the blanks.

Andrew Ng offers a highly practical analogy: Think of AI as a very smart, highly motivated recent graduate who just doesn't know you yet. If you simply tell them, "Please write a good self-evaluation for my boss," they have no idea what you actually accomplished over the past year and can only produce something generic and unhelpful.

What Power Users Do: Provide Sufficient Background

Ng believes AI power users almost develop a kind of "empathy" toward AI. While you shouldn't over-anthropomorphize it, you can put yourself in the position of whoever is receiving your instructions and ask: does this person actually have enough information to complete the task?

So power users will upload a wealth of information — screenshots of project trackers, recent project documents, even voice memos narrating project progress — before asking AI to write a self-evaluation. The result actually captures your proudest achievements.

Watch Out for Sycophancy: How to Get Honest Feedback from AI

One major issue with AI is that it often tries to please you. In fact, many AI systems are trained to maximize user satisfaction. If you ask a biased question, AI will often give you an answer that confirms your bias.

This "sycophantic" tendency isn't accidental — it stems from the training process. After pretraining, most mainstream large models go through RLHF (Reinforcement Learning from Human Feedback): human annotators rank multiple model responses, and the model learns "what kinds of answers humans prefer." The problem is that annotators tend to give higher scores to responses that feel pleasant and affirming, inadvertently training models to be agreeable. Anthropic's research team systematically named this phenomenon sycophancy and found that even when explicitly instructed to "be honest," models still show a significant tendency to accommodate users who express strong opinions.

For example, if you say "I have a great business idea — a mobile tie-dye service. Criticize it" — because you called it "great" and it's your idea, the AI will naturally want to please you and respond with "What a great idea!" This phenomenon is called sycophancy.

The Fix: Neutral Questions and Evaluation Rubrics

Andrew Ng offers two solutions:

First, ask neutral questions — don't reveal any preference for or against a particular answer.

Second, give AI a rubric or evaluation criteria to force objectivity. For example: "Please objectively analyze the following business idea — a mobile tie-dye service. Do not make things up. Use this rubric: Does it solve a real problem? Is there a market? Do I have a competitive advantage?"

Both methods work around the model's built-in people-pleasing tendencies at the prompt level. This way, AI doesn't know whether you want praise or pushback, and it's far more likely to deliver an honest assessment — like "This idea scores only 8 out of 100" — with specific reasoning.

Writing: Treat AI as a Thinking Partner, Not a Typing Machine

Andrew Ng observes that beginners and power users approach AI-assisted writing in completely different ways. A beginner will say "Write me a blog post about BlackBerry phones" and receive a wall of text that looks like "AI slop" — verbose, dull, and space-filling.

Power users don't ask AI to write directly. Instead, they use an iterative, step-by-step process:

1. First ask AI to generate an outline based on your uploaded notes
2. Provide feedback on the outline, revising back and forth until satisfied
3. Then expand the outline into bullet points, continuing to critique and iterate
4. Only then ask AI to develop it into full prose

This step-by-step approach is known in prompt engineering as "Prompt Chaining" or "Decomposed Prompting." The core idea is to break a complex task into simpler subtasks, where each subtask's output becomes the input for the next step. Research shows this approach not only significantly improves output quality but also keeps users in control and exercising judgment at every stage. Related techniques include "Tree of Thoughts," which lets models explore and backtrack across multiple reasoning paths.

In this process, you're essentially treating AI as a thinking partner — helping you brainstorm and explore different writing directions. This human-AI collaborative writing model fundamentally positions AI as a "first draft generator" and "idea stimulator" rather than the final author, with humans retaining editorial control and decision-making authority. The quality of output far exceeds what you'd get from one-shot generation.

AI Makes Fewer Mistakes Than You Might Think

AI systems do make mistakes — but if you prompt them correctly, the error rate may be far lower than most people assume. Andrew Ng points out that today's models are significantly more reliable than earlier versions.

The phenomenon of AI generating false information is known academically as "hallucination" — where the model outputs factually incorrect content with high confidence. This stems from the fundamental nature of large language models: they are probabilistic next-token predictors and don't truly "understand" facts. However, hallucination rates have dropped substantially from GPT-3.5 to GPT-4o and Claude 3.5 Sonnet. OpenAI's SimpleQA benchmark shows GPT-4o's factual accuracy has improved by tens of percentage points compared to GPT-3.5. Web search integration (e.g., Perplexity, ChatGPT's Browse feature) and tool use (allowing models to call calculators, databases, and other external tools) also help address hallucination at the system level.

Some AI errors that went viral on social media led people to believe AI makes mistakes constantly. Like the famous "how many R's are in strawberry" (it thought there were only two), or the amusing example — "I want to watch my car — should I walk or drive there?" AI answered "walk," but then you'd have no way to drive the car you're supposed to be watching.

Andrew Ng emphasizes that these viral cases are not representative of AI's true capabilities. True power users understand that AI can deliver enormous value through deep research, analytical reports, interpreting personal health and fitness data, even building websites.

Prompting Is a Scarce Workplace Skill

Andrew Ng concludes that regardless of your role, the ability to prompt AI at an expert level is a highly scarce and sought-after career skill. Becoming an AI power user not only saves time and improves your professional and personal life — it helps you build things you never expected to be able to create.

This view is being validated by market data. According to LinkedIn's 2024 Global Skills Trends Report, job demand for AI prompting-related skills grew by over 300% year-over-year. The World Economic Forum's Future of Jobs Report 2025 also ranked AI literacy as one of the fastest-growing core skills over the next five years. The value of prompting ability isn't just reflected in dedicated "prompt engineer" roles (some Silicon Valley companies offer salaries of $150K–$300K for these positions) — it shows up in productivity gains across all knowledge workers: market analysts using AI to assist with data interpretation, product managers using AI to generate PRD drafts, lawyers using AI to review contract terms. This "AI-augmented worker" model is redefining competency standards across roles.

Understanding one foundational aspect of AI — where its knowledge actually comes from — will help you better judge when to trust its answers and when not to. Large language models' knowledge primarily comes from the massive internet text data used during pretraining, which means they may have blind spots around events after their training data cutoff, specialized knowledge in niche domains, and information that was never publicly published. Understanding this is also the starting point for using AI wisely.

Key Takeaways

• From questioner to orchestrator: AI power users don't just ask questions — they upload complete materials, orchestrate multi-step tasks, and give models sufficient reasoning time
• Context determines quality: Like working with a smart but unfamiliar new colleague, providing sufficient background is a prerequisite for high-quality output
• Actively counter sycophancy: Use neutral questions and evaluation rubrics to work around AI's built-in people-pleasing tendencies and get honest feedback
• Iteration beats one-shot generation: Use prompt chaining and iterative writing workflows to position AI as a thinking partner, not a ghostwriter
• Rational perspective on AI errors: Viral mistake cases don't reflect model capabilities — with proper prompting, AI is far more reliable than most people imagine
• Prompting ability is competitive advantage: AI literacy is becoming one of the scarcest workplace skills — mastering it means gaining a systematic edge in efficiency and decision quality