Andrew Ng’s AI Prompting Course: Deep Research, Writing, Data Analysis

Course Overview

The course consists of three modules. Module 1 covers three levels of information retrieval: pretrained knowledge, web search, and Deep Research. Module 2 focuses on AI as a thought partner for brainstorming, writing, and critique. Module 3 teaches working with multimedia, code, building applications, and data analysis.

Chapter 1: Information Retrieval Tiers

Pretrained Knowledge – Direct query, uses the model’s internal knowledge, response time in seconds, suitable for definitions, common sense, or simple summaries.

Web Search – Automatic or manual trigger, fetches a few webpages, takes a few seconds to tens of seconds, ideal for up‑to‑date facts like weather or local business ratings.

Deep Research – Must be explicitly triggered, aggregates dozens to hundreds of webpages, takes minutes to tens of minutes, best for complex, multi‑perspective questions such as the impact of weather on tourism.

Key point: you must first decide whether your question is worth waiting a few minutes for the AI.

What Deep Research does internally?

Deep Research follows an agentic loop where the model plans a research agenda, launches parallel web searches, reads and evaluates relevance, optionally refines keywords, and finally synthesizes a report with citations.

你的问题
   ↓
[制定研究计划]  ←── Many products let you review/modify this plan
   ↓
[并行发起 N 个网页搜索]   ←── Not one‑by‑one, but simultaneous
   ↓
[阅读 + 评估相关性]
   ↓
[决定要不要换关键词再搜一轮?]  ──→ 是 → 回到上一步
   ↓ 否
[综合所有页面，加引用，输出报告]

The loop embodies the core definition of an agentic AI: the model decides the next search, whether to continue, and when to stop, rather than following a fixed pipeline.

Chapter 2: Combating AI Slop with Progressive Outlining

OpenAI data shows that writing accounts for 24% of all ChatGPT tasks and is the scenario with the highest failure rate, dubbed “AI Slop”. The author lists six observable fingerprints of AI‑generated text:

Excessive use of em dashes.

Overuse of buzzwords like “nuanced” or “delve”.

Three‑point list patterns even when only two points exist.

Formulaic “not X, but Y” constructions.

Adjective‑heavy noun phrases.

Vague, grandiose statements that sound important but lack substance.

Side effect: humans start sounding like AI, e.g., the word “delve” spikes in podcasts after ChatGPT’s release.

The recommended workflow, called Progressive Outlining, separates evidence gathering from drafting:

Step 1  研究证据 ────────► 让 AI 搜索/查找正反两方证据
Step 2  生成 3 种大纲 ───► brainstorm，给你选项
Step 3  你给反馈，迭代 ──► 改结构、加历史类比、调整顺序
Step 4  扩成 bullet 点 ──► 还不要正文，先扩成要点
Step 5  对 bullet 再迭代
Step 6  最后才让它写正文

Comparing direct writing versus outlining shows a leverage difference of roughly 1× versus 100×, meaning small changes in the outline can dramatically reshape the final article.

Chapter 3: Building Apps and Data Analysis with Code Execution

Two practical themes are covered: using AI to create small applications and using AI for data analysis, both relying on a “code execution” tool.

Prompt‑driven app examples

Fireworks click simulator

25‑minute Pomodoro timer

AA‑style bill splitter

Weather‑based outfit selector

All share three criteria: clear task definition, no external data or accounts needed, and they can run independently for a short period.

Difficulty spectrum (easy → medium → hard) is determined by whether the app requires networking, persistence, or multi‑user collaboration.

Data analysis pipeline

The AI first receives a CSV, decides the analysis requires computation, invokes the code execution tool, writes and runs Python (pandas/matplotlib), and returns a report with charts, numbers, and insights.

你扔过来一个 sales.csv
    ↓
模型推理：这事得算，不能瞎说
    ↓
调用 code execution 工具
    ↓
现场写 Python（pandas / matplotlib...）
    ↓
真的跑一遍代码
    ↓
返回带图、带数字、带洞察的报告

Example: uploading a sales file, the model identifies that only four drinks have significant changes and visualizes them, a capability beyond a typical Excel pivot table.

Warning: while simple analyses are strong, complex multi‑step analyses can hallucinate; always verify key numbers.

Tool‑Selection Decision Tree

The combined logic from Chapters 1 and 3 can be expressed as a decision tree:

收到你的问题
    │
    ├── 是常识/定义吗？        → 用预训练知识，几秒返回
    │
    ├── 需要时效性单点信息？   → 调用 Web Search
    │
    ├── 需要多源综合判断？     → 调用 Deep Research（用户必须主动开）
    │
    └── 需要算数/画图/处理数据？→ 调用 Code Execution，写代码跑代码

This tree is the core insight the course wants learners to internalize: the real task is selecting the appropriate AI tool, not merely crafting a fancy prompt.

Key Takeaways

Deep Research acts as a junior researcher that runs errands, not as a slower search engine.

Outlines are the lever; the body is the load—always work from the lever.

When a problem involves calculation or visualization, trigger code execution instead of relying on plain text generation.