Beyond the chat window, in brief
Lesson 3 continues the track’s first arc, the four lessons on using generative AI well. It adapts the session called Beyond Chatbots from the same Harvard Kennedy School course, and it pays the promise lesson 2 left open: what changes when these systems can speak out loud. It also solves a mystery lesson 2 could not: how some assistants arrive already knowing their job before you type a word.
The lesson asks nothing technical. Its raw material is the same task-from-your-own-week that lesson 2 used, and its practice builds the assistant lesson 2’s job seeker deserves: one that knows your background, your target role, and your resume because you told it once.
The capability: after this lesson, you can write standing instructions once so an assistant carries them into every conversation, point it at your own documents so it answers from your material, test where a tailored assistant’s lane ends, tell the two everyday tailoring methods apart from the specialist one, and ask who tailored any assistant you meet.
What the lesson covers. First the mystery’s answer, the system prompt: instructions and context written once, in a special place, and silently attached to every conversation, “valid for every single interaction.” Then retrieval, RAG for short: your documents become a knowledge base, the system searches it on every question and quietly attaches the best passages to the prompt, which brings answers newer than the model’s training cutoff and fewer invented ones, fewer but not zero. The course’s admissions-office activity supplies the lane rule: a tailored assistant is a specialist, and its lane is exactly as wide as what you fed it. Fine-tuning is the third method, where developers train the model further on examples from their field; in the course’s words it is “typically harder and more expensive,” a job for specialist teams, while the first two are “something that most of us can do.” A deliberately dated tour of how the course built assistants in 2024, with custom GPTs and Microsoft’s Copilot, treats those products as historical markers of a pattern that outlived them while they kept changing. The tailor warning follows: “the designer has a lot of control on how it’s going to respond,” so ask who wrote the instructions and chose the documents. The lesson closes with the machine’s new doors, sight, plain-English data analysis, and voice, held together by the course’s reminder that “the logic behind the tools is the same.”
Why this order. Lesson 2 taught you to assemble task, instructions, and context by hand, one conversation at a time; this lesson makes that anatomy permanent and industrializes the context, so the track’s prediction-machine model keeps explaining everything new. The practice in Clawless builds the two-capability assistant step by step and finishes with the lane-edge test. Lesson 4 turns capability into judgment: should AI do this task at all, the course’s decision framework for that question, and a cautionary tale about an official chatbot that gave the public wrong answers.