Summary: Beyond the chat window, tailoring AI to your work

All week you have been opening fresh chat windows and re-introducing yourself: who you are, what the project is, how you like things written. Then a coworker asks her assistant what the travel policy says about short-notice flights and gets the answer in seconds, in plain company language, pointing at the actual policy. No setup. Her assistant arrived already knowing its job. This lesson explains how, with two ideas you can use this week, and raises the quieter question of who put that knowledge there.

Core ideas

• A system prompt is instructions and context written once, in a special place, and silently attached to every conversation that follows: what the tool should do and not do, how it should respond, what it should know about you. Whatever you put there stays “valid for every single interaction.”
• The course demonstrates with a silly standing instruction, answer everything in rhyme, then a practical one: my time is scarce, keep every reply under a hundred words in crisp bullet points. Every future conversation obeys, no reminders.
• Nothing mystical happened. The machine is still a prediction machine writing one word at a time from everything in front of it; a system prompt guarantees certain words are always in front of it. An assistant that arrives knowing its job is one whose job description was quietly stapled to page one.
• A system prompt should hold only what is stable: who you are, how you like output, the job. Per-task detail belongs in the task’s prompt.
• Retrieval, formally retrieval augmented generation or RAG, sounds harder than the idea deserves. You give the AI documents, websites, or databases that matter to you, a pile the course calls a knowledge base. Ask a question and the system searches that knowledge base, quietly attaches the most related passages to your prompt, and answers from the combined text.
• Retrieval is not a new kind of machine. It is the same prediction machine with an automated librarian standing in front of it. Lesson 2 called context the part almost everyone skips and the part that helps most; retrieval is context, industrialized.
• Two benefits: your documents can be newer than the model’s training cutoff, and grounding the machine in real pages can help cut down on confidently made-up answers. Cut down, not eliminate. A grounded assistant can still misread its ground.
• The course’s students built an admissions assistant from standing instructions plus the school’s public admissions website. Good at: pulling answers out of its material in a friendly, professional tone, something ordinary software found very hard before generative AI. Bad at: anything its documents do not cover. A tailored assistant is a specialist. Its lane is exactly as wide as what you fed it.
• The third method, fine-tuning, has developers train the model further on examples from their field, so the model itself changes. In the course’s words it is “typically harder and more expensive,” a job for specialist teams. The first two methods are “something that most of us can do.”
• How the course built these in 2024: a ChatGPT feature called custom GPTs (standing instructions in a form, uploaded files, a shareable assistant, behind a subscription) and Microsoft’s Copilot, which turned a single document into a slide deck about twenty-five slides long. Treat the names as historical markers; the products kept changing while the pattern spread. What survived is standing instructions, retrieval, and the discovery that ordinary people can tailor an assistant without code. As of mid-2026, AI features sit inside much of the software people already use.
• The warning: “the designer has a lot of control on how it’s going to respond.” When the designer is you, that control is the point. When it is anyone else, the assistant’s helpfulness is shaped by choices you cannot see. Nothing sinister is required; a store’s assistant is tuned to sell. Every tailored assistant has a tailor. Ask who.
• Sight: the course’s instructor drags in a photo of his fridge and gets a dinner proposal read off the shelves. The fridge is a party trick; the capability is not. As of mid-2026, most major assistants can look at an image you give them, though what any tool does well with images this month is a moving target.
• Data: upload a spreadsheet, ask in plain English for a chart of average income by education level, and back one comes, no formulas. Not how you would do serious analysis, but a question that once needed a specialist is now a sentence.
• Voice: as of mid-2026, many assistants can hold a spoken conversation at nearly the pace of a phone call. In most tools the mechanics are humble: speech becomes text, the prediction machine does its usual work, and the answer returns as sound. Same machine, new door.
• Underneath all of it, “the logic behind the tools is the same”: prediction over an enormous body of human writing, tuned to be helpful. This lesson gave the machine a job description, a filing cabinet, eyes, and a voice. The trick did not change, and it is still learnable.

What changes for you

The re-pasting stops today: anything you type into every conversation is a system prompt waiting to be written. Your documents become your edge, because a model grounded in your files can finally be specific about your world instead of generically helpful about everyone’s. And every assistant you meet now comes with a label you know how to read: instructions you cannot see, documents you did not choose, a lane, and a tailor. The next lesson turns capability into judgment: should AI do this task at all, with the course’s decision framework and a cautionary tale about an official chatbot that gave the public wrong answers.