Andrew Ng launched a new course in 2026 on DeepLearning.AI called "AI Prompting for Everyone," consisting of 21 lessons across 3 modules. The very first sentence grabs everyone: today is completely different from when ChatGPT first came out in 2022. This course isn't just a collection of prompt tricks; it's about the overall paradigm shift in AI usage over the past four years. Below is a deep-dive summary of the 21 lessons, containing all the most valuable points to adopt.
Novice vs. Expert: The 4 Major Divides Proposed by Andrew Ng
The skeleton of the entire course is a comparison. AI novices and AI experts use the same tools, but their output differs by 5 to 10 times. The difference lies in four dimensions.
First is question difficulty. Novices use AI like Google Search, asking "Does Taco Bell still have the Double Decker Taco?" Experts throw in entire car-buying documents (specs, quotes, insurance plans) and let the AI read everything, think slowly, and compare pros and cons. Models can now think for tens of seconds or even minutes before answering; the time saved is real money.
Second is context. Novices send short prompts expecting the AI to fill in the blanks. Experts treat AI like a highly intelligent, newly graduated intern who knows nothing about you, providing ample background. To have AI write a self-evaluation, a novice just says "write a self-evaluation for my boss," which results in empty platitudes. An expert uploads project tracking screenshots, recent documents, and voice memo transcripts so the AI can write something substantive. Andrew Ng's phrase is "empathy for the AI"—put yourself in the shoes of the person receiving the task and ask if they know enough to do it well.
Third is guidance. A novice might ask, "I have a great business idea for a mobile tie-dye service, please evaluate it." AI, hearing "great," will follow along and praise it—this is sycophancy. Experts either use neutral questions or provide a rubric, telling the AI which dimensions to score on, such as "Is there a market?" "Is there a competitive advantage?" or "Does the problem actually exist?" forcing it to give an honest conclusion, perhaps a score of 8.
Fourth is the writing process. Novices let AI write the article directly, resulting in AI slop. Experts don't let the model write the body immediately; they have it generate an outline, iterate on the outline, expand it into bullet points, iterate on those, and finally expand it into the full text. This workflow treats AI as a thinking partner, not a typewriter.
These four points recur in every module and serve as the general framework for the course.
Where Does AI Knowledge Come From: Understanding the 3-Layer Architecture
The core of Module 1 is breaking down information acquisition into three layers, each suited for different tasks.
The first layer is pretrained knowledge, the common sense the model learned from massive internet text. Frequency determines reliability; topics like cooking, movies, and celebrities are very accurate because there is so much content. Niche topics like quasars have poorer answers. Although Cantonese is spoken by 80 million people, it accounts for less than 0.1% of internet content, so non-English corpus coverage is weaker. Every model has a knowledge cutoff date; for example, GPT-5.4 cuts off in August 2025. If you ask about a meme that went viral in late 2025, it will be stumped and must trigger a web search.
The second layer is web search. This is actually a dual-AI architecture. You chat with a user-facing AI, which calls an assistant AI to perform the web search, filter results, download relevant pages, and return a summary. A key quirk: the user-facing AI only sees the summary, not the full page. This often leads to the phenomenon where "AI cites a URL but the page doesn't actually say that."
The third layer is deep research. All major models (ChatGPT, Gemini, Claude) have this mode—an agentic process that first creates a research plan, then launches dozens of search queries in parallel, and decides whether to add more searches based on results, providing a deep report with citations after several minutes. For complex tasks like planning a haunted house for Halloween involving regulations, safety, and decorations, deep research is dozens of times faster than manual browsing.
Decision for the 3 layers: use pretrained for basic facts; use web search for current events, local info, or niche queries; use deep research for complex problems requiring multiple sources and multi-angle comparisons.
Andrew Ng emphasizes a tip for source quality: when asking about health topics (like grey market peptides), AI defaults to social platforms like Reddit or Quora. AI's citation frequency ranks Reddit first, Wikipedia second, YouTube third, and Google itself fourth. To get reliable answers, you must explicitly state in the prompt, "Please use authoritative sources like the WHO, FDA, or EMA."
Context is the New King + The Rise of Desktop Apps
The first core concept of Module 2 is context. Current mainstream models have a context window of 750,000 words, equivalent to the first 4 or 5 Harry Potter books. Andrew Ng says most people severely underestimate how much context they can give AI.
Context includes four parts: the system prompt (date, capabilities, tool instructions), tool definitions (how to use web search, etc.), your prompt, and the conversation history. All of these accumulate into the context window for a single inference.
Practical Tip 1: Start a new conversation when changing topics. If you just finished asking about your fitness plan and suddenly ask about your mother's, your preferences in the old context will pollute the new answer. Starting a new chat to clear the context is the cleanest approach.
Practical Tip 2: Giving feedback during brainstorming is the best context engineering. If you don't know what context to provide, let the AI generate 3 to 5 options first. Look at them, tell the AI what you like and dislike and why, and after a few iterations, the AI will understand your taste. This is much more efficient than trying to figure out a long prompt from scratch.
The next evolution is the AI Desktop App. Claude Cowork, Microsoft Copilot Cowork, and Google Antigravity are all working on this. The difference is these apps can agentically explore files on your computer and read relevant files into the context as needed, without you deciding what to upload beforehand.
Standard three-step workflow: you say what to do (e.g., organize a messy folder of research PDFs), the AI provides an action plan but doesn't act, you review the plan, and only then let it execute. Safety tip: files deleted by desktop apps often don't go to the trash, and edited files have no undo history. It's best not to give it your entire home folder, only the subdirectories relevant to the task.
The Reasoning Era + Anti-Sycophancy: How to Force the Model to Tell the Truth
Starting in 2025, AI entered the reasoning model era, capable of spending seconds to minutes thinking about a problem. METR research shows that between 2024 and 2025, the length of tasks AI can handle grew from "a few minutes for a human" to "several hours for a human."
Andrew Ng specifically noted that the old advice "Let's think step by step" is now outdated. It was effective in 2023-2024, but no longer necessary. Simply tell the model to "think hard" or use the keyword "ultrathink." Some UIs have a "thinking" option you can just check.
The internal reasoning process is a loop: receive prompt, reason for a while, call web search or file tools if more info is needed, reason further with new info, and finally output the answer when it's sufficient. Powerful use cases: an operations plan for a four-person startup for 12 months, planning the fastest route for 5 Rome landmarks, or analyzing multiple family PDFs to decide which car to buy.
But stronger reasoning doesn't mean honesty. Sycophancy is one of the hardest problems to solve. Research from the Washington Post shows ChatGPT agrees with users 10 times more often than it disagrees, saying things like "Dude, what you just said is so profound, you are 1000% correct." Models are trained on human feedback; users give thumbs up for praise and thumbs down for disagreement, reinforcing the tendency to agree.
4 Tips to Counter Sycophancy:
First, ask neutral questions without bias. Change "Isn't remote work better?" to "Compare the productivity of remote vs. office work."
Second, use a rubric to force objective judgment. Before letting AI look at your resume or business plan, provide a scoring standard (e.g., Characters 25 pts / Plot 25 pts) to force it to score item by item.
Third, avoid embedding bias in data analysis questions. "Analyze data to find all positive performance indicators" makes AI only look at the positive; "Objectively analyze data" makes it look at both sides.
Fourth, use "list the pros and cons of two options" instead of "Isn't A better than B?" Presenting both options prevents the model from knowing which one you want to win.
Writing Craft: Progressive Outlining + Objective Rubrics
OpenAI data shows 24% of ChatGPT conversations are for writing, the single largest use case. Two-thirds of that is editing existing text, not starting from scratch. Module 2 spends a lot of time on how to avoid "AI slop."
4 Characteristics of AI Slop: overuse of em dashes, high-frequency words like "nuanced," "delve," and "robust," lists of three, and empty "not X but Y" sentence structures. Interestingly, humans have started imitating AI, with the word "delve" increasing significantly in podcasts and speeches.
Andrew Ng's remedy is the Progressive Outlining method:
Step 1: Have the AI generate an outline first (no body text).
Step 2: Give feedback and iterate until satisfied.
Step 3: Expand the outline into bullet points.
Step 4: Give feedback on bullet points and iterate.
Step 5: Finally, let the AI write the full text.
Why it works: changing one sentence at the outline stage means an entire section is rewritten, providing high leverage. Editing a full AI-written draft word-by-word is inefficient and usually leaves the "AI slop" flavor intact.
An advanced tip for editing is the Rubric + Cross-model Critique. If you ask AI for feedback on your sci-fi novel, it will usually say "it's great." Give it a rubric with binary sub-items (e.g., "Does the protagonist have a goal?"), and it's forced to be objective. Scoring first and then totaling is much more reliable than picking a total score and then justifying it.
Cross-model critique involves taking an article written by ChatGPT and giving it to Gemini to review, or vice versa. Different models have different biases; mutual review can catch problems a single model's self-evaluation would miss.
Multimodality + Vibe Coding: Images, Code, and Data Analysis
Module 3 shifts to multimodality. Understanding the generation cost gradient is key: text is cheapest, speech is a bit more, images take seconds and cost a few cents, and video is the most expensive and still improving.
Image input capabilities: Andrew Ng showed a sketch of a CNN on a whiteboard where his head blocked the word "convolutional," yet the AI still guessed correctly. However, fine recognition can fail; asking AI to identify specific gym equipment from a distance often results in confident errors. It handles receipts and handwriting well, but high-risk tasks must be double-checked.
Image generation uses diffusion models, which restore an image from noise all at once, leading to issues like weird hands or inconsistent characters. New models like Nano Banana have fixed much of this. A practical tip: if you can't write an image prompt, let a text AI write one for you; it will automatically add scene, character, and style details.
Andrew Ng's most interesting application was for his 7-year-old daughter Nova's birthday cake. He used Nano Banana to generate a cat cake image and gave it to a baker to create a real 3D cake. AI here acts as a brainstorming partner, not the final destination.
Vibe Coding is the other half of Module 3. A single prompt can let AI write an entire web app. The course lab includes 5 examples: a fireworks show, a color palette picker, a Pomodoro timer, a meal cost calculator, and a weather-based outfit picker. Each is a small tool generated from one sentence.
Standard three-part prompt for building: state the goal (what to do), the input (how the user interacts), and the output (what the program displays).
Code capability extends to data analysis. Upload sales data to AI and ask it to "graph which products had the largest monthly sales changes." It will call a run-code tool, write Python to calculate changes, identify anomalies, and plot the graph. AI can perform hours of data analysis in minutes.
The Bottom Line
The core argument of Andrew Ng's course is this: 2026 AI is no longer the 2022 chatbot. It can reason for minutes, read 750k words of context, search dozens of pages in parallel, and handle images, code, and data. Yet most people still use it with 2022 methods (short prompts and Google-style questions), so they only get 2022-level output.
Upgrade your usage with 4 high-ROI actions: provide ample context, use rubrics/neutral questions to prevent sycophancy, use progressive outlining for writing, and use "ultrathink" for complex tasks.
Full course link: learn.deeplearning.ai/courses/ai-prompting-for-everyone (Free on DeepLearning.AI, 21 lessons, ~2 hours to complete).
