Beyond the chat window: tailoring AI to your work

All week you have been opening fresh chat windows and introducing yourself. Who you are, what the project is, how you like things written. Paste, ask, repeat. The answers are good, lesson 2 saw to that. But by Thursday you feel less like a person with an assistant and more like a receptionist for one. Then a coworker stops you cold. She opens her assistant and simply asks what the travel policy says about booking flights on short notice. The answer arrives in seconds, in plain company language, pointing at the exact section of the actual policy. No setup. No pasted paragraphs. Her assistant arrived already knowing its job.

Two questions should be forming. First, how? Nothing in lesson 2 explains an assistant that knows things before the conversation starts. Second, and quieter: who put that knowledge there, and what else did they put in? This lesson answers both. The how is two simple ideas you can use yourself this week, no programming required. The who is one of the most practical warnings in this track. And behind both sits a bigger shift. These systems no longer need the chat window at all, because they can see your images and speak out loud.

This lesson adapts the session called Beyond Chatbots from the Harvard Kennedy School course this track is built on, created by Sharad Goel, Dan Levy, and Teddy Svoronos. Dan Levy teaches this one, and his delight keeps breaking through. One demo made his jaw drop. Another left him openly astonished. By the end you will be able to build the modern version of what impressed him.

Instructions that do not wash away

Start with the coworker’s mystery. The trick has a name: the system prompt.

Think back to lesson 2: task, instructions, context. Your instructions and context have an annoying property. They wash away. Close the chat, and the machine’s next conversation ordinarily starts blank. That is exactly why you have been re-pasting those paragraphs all week.

A system prompt fixes that: 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. As Levy tells the class, whatever you put there stays “valid for every single interaction.” You type it one time. The machine reads it every time.

The course demonstrates with a deliberately silly standing instruction: tell the assistant once to answer everything in rhyme, and every future answer rhymes. Then the practical one: tell it your time is scarce and every reply should stay under a hundred words, in crisp bullet points. Every reply in every new conversation now arrives short, crisp, and bulleted, no reminders.

Hold this against the track’s core insight. The machine is still a prediction machine writing one word at a time from everything in front of it. A system prompt simply guarantees that certain words are always in front of it, before you type a thing. An assistant that arrives already knowing its job is an assistant whose job description was quietly stapled to page one.

Point it at your own documents

A system prompt can tell the machine how to behave, but it cannot hand it your company’s two-hundred-page travel policy. For that there is the second method, retrieval.

The formal name, retrieval augmented generation, RAG for short, sounds far harder than the idea deserves. You give the AI documents, websites, or databases that matter to you. The course calls this pile a knowledge base. Ask a question and the system first searches that knowledge base for the most related passages. It quietly attaches what it found to your prompt, and the machine answers from the combined text. You never do the retrieval yourself. You just see an answer grounded in your material. Notice this is not a new kind of machine. It is the same prediction machine with an automated librarian standing in front of it. In lesson 2, context was the part of the prompt almost everyone skips and the part that helps most. Retrieval is context, industrialized: the right pages, fetched for you, on every question.

Two benefits follow. The model’s training data stops at a cutoff date. Your documents can be newer than that. And grounding the machine in real pages can help cut down on confidently made-up answers. That failure mode gets its own treatment near the track’s end. Notice the hedge: cut down, not eliminate. A grounded assistant can still misread its ground.

The course then hands students the whole toolkit: build an assistant for the Kennedy School’s admissions office. They gave it standing instructions and pointed it at the school’s public admissions website. Most rated their result good or very good. Then the more instructive question: what is it good at, and what is it bad at? Good at: pulling answers out of its material in a friendly, professional tone. Ordinary software found that very hard before generative AI. Bad at: anything its documents do not cover. The bot knew only the public website, and no politeness saves it past that edge. A tailored assistant is a specialist. Its lane is exactly as wide as what you fed it.

There is a third method, fine-tuning. Developers train the model further on examples from their field. The model itself changes, not just what it reads. That is how you get an assistant genuinely steeped in medical research. It is also, in the course’s words, “typically harder and more expensive,” a job for specialist teams. The first two methods are, as Levy puts it, “something that most of us can do.” That sentence is the practical heart of this lesson.

Build one yourself, and read the label on everyone else’s

A note on how the course did this in 2024, because those products kept changing while the pattern spread. When the class was filmed, the showcase way to build one was a ChatGPT feature called custom GPTs. You typed standing instructions into a form, uploaded files, and got a shareable assistant, if you paid for a subscription. The season’s other star was Microsoft’s Copilot, AI folded into office software. Levy watched it turn a single document into a slide deck about twenty-five slides long, presenter notes included, and his jaw dropped. Treat those names the way lesson 1 taught you to treat the 2022 launch date: as historical markers. The products have shifted many times since. What survived is everything this lesson teaches: standing instructions, retrieval, and the discovery that ordinary people can tailor an assistant without code. The direction survived too. As of mid-2026, AI features sit inside much of the software people already use. The chat window is no longer the only place you will meet this machine.

Now the modern version. Clawless, the working environment we use across Clawdemy, supports both moves you just learned. You can give an agent standing instructions that persist, its job description written once. You can point it at your own documents, so it works from your material. This lesson’s practice builds the assistant lesson 2’s job seeker deserves. It knows your background, your target role, and your resume before you type a word, because you told it once.

Before you build, settle the hook’s second question. The coworker’s assistant knew things because someone wrote its instructions and chose its documents. That someone had goals. The course puts the warning plainly: “the designer has a lot of control on how it’s going to respond.” When the designer is you, that control is the whole point. When it is anyone else, the assistant’s helpfulness is shaped by choices you cannot see: what it emphasizes, what it plays down, what it never brings up. Nothing sinister is required. A store’s assistant is tuned to sell. Every tailored assistant has a tailor. Get in the habit of asking who.

When the machine can see and speak

Everything so far tailored what the machine knows. The session’s final stretch changes what it can take in, and here Levy’s composure gives out entirely: “the first time I saw these features I was blown away.”

First, sight. Levy drags a photo of a dinner plate into the chat, salmon and asparagus, and asks for a recipe. The machine identifies the dish and writes a plausible one. Then the better demo: a photo of the inside of his fridge, and a request for dinner based on what is in it. The machine reads the shelves and proposes stuffed bell peppers with a side salad. 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. People lean on that daily: error messages, forms in other languages, photos of whatever needs explaining. What any tool does well with images this month is a moving target. That it can look at all is now the norm.

The same session shows him uploading a spreadsheet and asking, in plain English, for a chart of average income by education level. Back one comes, no formulas. Not how you would do serious analysis, Levy is careful to say, but a question that once needed a specialist is now a sentence.

And voice, which lesson 2 promised you. As of mid-2026, many assistants can hold a spoken conversation: you talk, it talks back, at nearly the pace of a phone call. In most tools the mechanics are humble. Speech becomes text, the prediction machine does exactly what lesson 1 said it does, and the answer returns as sound. Same machine, new door. For anyone whose hands are full, or who simply thinks out loud, the chat window stops being a requirement at all.

Levy tells the class up front that even as the tools multiply, “the logic behind the tools is the same.” He means the machinery you already understand: prediction over an enormous body of human writing, tuned to be helpful. Nothing in this lesson replaced the prediction machine that writes one word at a time. We gave it a job description, a filing cabinet, eyes, and a voice. It is still the same trick, and the trick is still learnable.

Why this matters when you use AI

The re-pasting stops today. Anything you find yourself typing into every conversation is a system prompt waiting to be written. Promote it once and the setup tax disappears.
Your documents are your edge. A general model knows the internet’s averages. Grounded in your files, your policies, your notes, your resume, it can finally be specific about your world instead of generically helpful about everyone’s.
You can now read the label. Every assistant you meet, at work, in a support chat, inside your software, is a tailored assistant: instructions you cannot see, documents you did not choose. Knowing the anatomy tells you what to trust it on (its lane) and what to ask about (its tailor).

Common pitfalls

Trusting a specialist outside its lane. The admissions bot was excellent about its own website and lost beyond it. Ask what an assistant was given before you rely on it. Its confidence will not change at the lane’s edge, but its accuracy will.

Stuffing the standing instructions with everything. A system prompt applies to every conversation, so it should hold only what is stable: who you are, how you like output, the job. Per-task detail belongs in the task’s prompt.

Forgetting the tailor. An assistant that arrives knowing its job learned that job from someone. When that someone is not you, their goals ride along in every helpful, friendly answer. Warmth is part of the design.

Expecting retrieval to end wrong answers. Grounding reduces invention; it does not guarantee comprehension. The machine can cite your document and still misread it. For answers that matter, follow the reference back to the page.

What you should remember

A system prompt is standing instructions and context, written once and attached to every conversation. It is lesson 2’s anatomy made permanent, and the reason some assistants arrive already knowing their job.
Retrieval, RAG for short, searches your own documents and slips the best passages into the prompt behind the scenes. Fresher answers, more specific answers, fewer invented ones. Fewer, not zero.
Fine-tuning trains the model further, changing the model itself, and belongs to specialist teams; the other two methods are within almost anyone’s reach.
The 2024 products that demonstrated all this are dated markers; the tailoring pattern outlived them, and your practice runs it today in Clawless.
Vision, voice, and plain-English data analysis are new doors into the same machine. The logic underneath never changed: prediction, one word at a time.

What’s next

You now have real capability: prompts, tailored assistants, a machine that can look and listen. The next skill is judgment. Should AI do this task at all? Lesson 4 gives you the course’s decision framework for that question, plus a cautionary tale about an official chatbot that answered the public confidently and wrongly. Bring a task you are tempted to hand over; you will run it through the filters yourself.

If you remember one thing

A tailored assistant is not a smarter machine. It is the same machine, better briefed.
Standing instructions give it its job. Your documents give it your world.
When the tailor is you, that is the power. When it is someone else, that is the question.