Andrew Ng's Prompting Course: 4 Key Differences Between AI Beginners and Power Users

Why Prompting Has Become a Core Career Skill

The way we interact with AI models today looks nothing like it did when ChatGPT first launched. In his latest AI prompting course, Andrew Ng makes one thing clear from the start: using AI effectively is one of the most impactful skills you can develop.

Prompt engineering emerged as a distinct discipline after ChatGPT's release in 2022. It refers to the systematic approach of crafting input instructions to optimize the output quality of large language models. Early practitioners were reportedly earning over $300,000 annually, drawing widespread attention. As model capabilities improved, debate grew over whether "prompt engineer" should be a standalone role — the more mainstream view now holds that prompting should be a foundational skill for all knowledge workers, not a niche specialty. Ng's course was built precisely with this shift in mind.

The course's central insight: today's AI tools are far more powerful than they were a year ago, yet many people who haven't stayed on the cutting edge remain frustrated by disappointing outputs. The problem usually isn't the tool — it's the user. By contrasting the experiences of AI beginners and power users, Ng reveals four key dimensions that separate them.

Difference #1: Give AI Complex Tasks — Don't Treat It Like a Search Engine

The most common mistake AI beginners make is using AI like Google — asking simple questions that can be answered instantly, while missing where it actually shines.

Ng emphasizes that AI's real strength lies in handling complex, high-difficulty tasks. He uses buying a car as an example: you can upload vehicle specs, quotes, and insurance documents to ChatGPT, Gemini, or Claude, then ask it to analyze the tradeoffs between several models.

These kinds of tasks prompt the AI to think deeply for tens of seconds or even minutes, ultimately producing a thorough analytical report. The key mindset shift: don't be afraid to give AI complex tasks that require deep thinking, and give it enough time to complete them. That's where AI truly delivers as a productivity tool.

Difference #2: Think of AI as a Brilliant New Grad Who Doesn't Know You Yet

This is the most illuminating analogy in the entire course. Ng suggests thinking of AI as a very smart, highly motivated recent graduate from a top school — but one who knows absolutely nothing about your background or needs.

Beginners tend to use a single terse prompt and expect the AI to fill in the blanks. Saying "write my annual self-evaluation for my boss" gives the AI no information about what you actually accomplished — so it produces a generic, one-size-fits-all template.

Ng introduces a key concept here: empathy toward AI. He suggests asking yourself: if a person received exactly this set of instructions, would they have enough information to do an excellent job?

This is grounded in real technical reality. Modern LLMs have a "context window" — the maximum amount of information they can process in a single conversation. GPT-4 Turbo supports 128K tokens; Claude 3 supports 200K tokens. A larger context window means you can upload complete project documents, contracts, and data sheets, allowing the AI to generate responses with a thorough understanding of the situation — the more complete the information the model can "see," the higher the quality of its reasoning.

Power users provide rich context: screenshots of project trackers, recent project documents, even voice memos narrating recent progress — before asking the AI to write the self-evaluation. The result actually reflects their most meaningful achievements. The richness of context directly determines the ceiling on output quality.

Difference #3: See Through the "Sycophancy" Trap — Use a Rubric to Force Honest Feedback

This section exposes one of AI's lesser-known but genuinely dangerous traits: sycophancy.

Ng points out that many AI systems are trained to "please users." Sycophancy stems from the RLHF (Reinforcement Learning from Human Feedback) training mechanism: models are refined through human rater feedback, and humans naturally give higher scores to responses that feel agreeable. This teaches the model to "please" users rather than "help" them. In 2023, Anthropic, OpenAI, and other organizations published research confirming that sycophancy is widespread among mainstream models, identifying it as a core challenge in AI alignment.

If you ask a leading question, the AI will often give you the answer you seem to want. Say "I have a brilliant business idea — mobile tie-dye services. Critique it for me." Because you used the word "brilliant," the AI is naturally inclined to flatter you with "What a great idea!" The moment you hint at your preferred answer, the AI is very likely to mirror your bias right back at you.

Ng offers two ways to break this pattern:

Use Neutral Questions

Don't give any hints about the answer you're hoping for. Keep questions objective and neutral so the AI has no basis for guessing your stance.

Provide a Rubric

A more advanced approach is to give the AI an explicit evaluation rubric. For example: "Please objectively analyze the following business idea — a mobile tie-dye service — and score it on these criteria: Is there a real pain point? Is there market demand? Do I have a competitive advantage?"

When you frame it this way, the AI can't tell whether you want praise or criticism, so it's far more likely to give an honest assessment — something like "I'd give this idea a 6 out of 10" with clear explanations for the deductions. A rubric is one of the most effective tools for forcing AI to stay objective. Researchers are also exploring systemic approaches like Constitutional AI and adversarial training to reduce sycophancy at the technical level — but until those mature, prompt-side strategies remain the most immediately accessible defense.

Difference #4: Use an Iterative Workflow to Escape "AI Slop"

Beginners and power users approach writing tasks in completely different ways. A beginner says "write a blog post about BlackBerry phones" and gets back a pile of text that sounds like "AI slop" — lots of words, no real substance.

Ng recommends a step-by-step iterative workflow rather than asking the AI to produce a finished piece in one shot. This aligns closely with the research concept of Chain-of-Thought Prompting — a 2022 paper from Google Brain systematically demonstrated that guiding a model to reason step by step rather than jump to an answer significantly improves accuracy on complex tasks. A multi-step workflow applies this principle to content creation, with each step serving as an independent reasoning task that constrains and validates the next:

1. Start with an outline: Based on uploaded notes, ask the AI to "generate a blog outline from my notes";
2. Critique and iterate the outline: Give feedback on the outline and refine it back and forth until you're satisfied;
3. Expand into detailed bullet points: Flesh out the outline into detailed points, and optimize again;
4. Write the final draft last: Only once you're happy with the skeleton do you ask the AI to expand it into the final text.

The essence of this workflow is treating AI as a thinking partner — helping you explore different directions and refine your structure — rather than a one-click drafting machine. It's the creative equivalent of agile development: get feedback at every checkpoint, rather than discovering you've gone off course only at the end. The content produced through iterative refinement far surpasses anything generated by a single-shot "AI slop" prompt.

AI's Real Capabilities Far Exceed What Viral Failure Cases Suggest

At the end of the course, Ng sets the record straight on AI's track record. He acknowledges that AI does make mistakes — but likely far less often than most people assume, especially when you prompt it well.

Early AI models made far more errors than today's. But certain viral mistakes — like the "how many R's are in 'strawberry'" fail or the infamous "I'm going to see my car, should I walk or drive?" answer — have caused people to dramatically overestimate AI's current error rate. It's worth noting that these failures often relate to how LLMs process text: models operate on tokens, not individual characters, which explains why early models stumbled on seemingly trivial tasks like letter counting. As model architectures have evolved and targeted training has improved, these basic errors have dropped significantly — but public perception tends to lag behind technical progress.

Ng is clear: these viral examples are not representative of AI's true capabilities. Power users understand that AI can deliver enormous value on tasks like deep research, writing analytical reports, interpreting health data, and even building websites.

The ability to prompt at an expert level is a highly sought-after career skill, regardless of your field. Understanding where AI's knowledge comes from and where its limits lie is the foundational mindset for making the most of it — and that's exactly where the prompt engineering learning journey begins.

Key Takeaways