Practice: Beyond the chat window, tailoring AI to your work
Self-check
Section titled “Self-check”Seven short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. What is a system prompt, and what problem does it fix?
Show answer
Instructions and context written once, in a special place, and silently attached to every conversation that follows. It fixes the washing-away problem: ordinarily your instructions and context vanish when the chat closes, and the machine’s next conversation starts blank. Whatever you put in a system prompt stays “valid for every single interaction.”
2. What belongs in standing instructions, and what belongs in the task’s prompt instead?
Show answer
Only what is stable belongs in the standing instructions: who you are, how you like output, the job. Per-task detail, this posting, this draft, this deadline, belongs in the task’s prompt. A system prompt applies to every conversation, so stuffing it with everything pollutes them all.
3. Walk through what happens when you ask a question to an assistant built with retrieval.
Show answer
The system first searches your knowledge base, the documents, websites, or databases you gave it, 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. It is the same prediction machine with an automated librarian standing in front of it.
4. What two benefits does retrieval bring, and what is the hedge on the second one?
Show answer
First, freshness: the model’s training data stops at a cutoff date, and your documents can be newer than that. Second, grounding the machine in real pages can help cut down on confidently made-up answers. The hedge: cut down, not eliminate. A grounded assistant can still misread its ground, so for answers that matter, follow the reference back to the page.
5. What was the course’s admissions assistant good at and bad at, and what rule does that give you?
Show answer
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 did not cover. It knew only the public admissions website, and no politeness saved it past that edge. The rule: a tailored assistant is a specialist, and its lane is exactly as wide as what you fed it.
6. What is fine-tuning, and who is it for?
Show answer
Developers train the model further on examples from their field, so the model itself changes, not just what it reads. That is how you get an assistant genuinely steeped in medical research. In the course’s words it is “typically harder and more expensive,” a job for specialist teams. The other two methods, system prompts and retrieval, are “something that most of us can do.”
7. Why should you ask who tailored an assistant?
Show answer
Because someone wrote its instructions and chose its documents, and that someone had goals. As the course puts it, “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. Nothing sinister is required. A store’s assistant is tuned to sell.
Try it yourself
Section titled “Try it yourself”Build the assistant lesson 2’s job seeker deserves
In lesson 2 you built a three-part prompt around a real task and re-typed the context every time. This practice makes that setup permanent. Clawless, the working environment we use across Clawdemy, supports both moves from the lesson: you can give an agent standing instructions that persist, and you can point it at your own documents. Open Clawless and build the job search assistant step by step.
- Draft the standing instructions before you type them anywhere. On paper or in a note, write three short blocks: who you are (your background in two or three sentences), what job this assistant has (helping you land a specific target role), and how you like output (length, tone, format, for example replies under two hundred words unless asked). Check each line against the stable-versus-per-task rule: if it would change from one task to the next, cut it. A specific job posting is per-task detail; your target role is stable.
- Give the agent those standing instructions. In Clawless, set your draft as the agent’s persistent standing instructions, its job description written once. From now on every conversation with this agent starts with that knowledge already in place.
- Point it at your documents. Gather your resume, plus one or two other pieces of your own material if you have them, such as notes on target companies or a past cover letter you liked. Give them to the agent so it works from your material. This pile is your knowledge base, and it is what makes the assistant specific about your world instead of generically helpful about everyone’s.
- Run the no-setup test. Start a fresh conversation and ask, with no preamble, what roles you are targeting and what your strongest qualifications are. A good answer here should draw on your standing instructions and your resume without you pasting a thing. That is the coworker’s trick from the lesson’s opening, reproduced: it knows your background, your target role, and your resume because you told it once.
- Put it to work on a real task. Paste a job posting into the conversation, fenced off the way lesson 2 taught, and ask the assistant to identify which of your resume’s experiences best match the posting and to draft a tailored opening paragraph. Notice the division of labor: the posting rode in with the task, everything else was already there.
- Run the lane-edge test. Now ask it something its documents do not cover, for example what salary the posted role pays, or what the hiring manager is like. Watch the tone: its confidence will not change at the lane’s edge, but its accuracy will. Anything it offers here is coming from the general model, not from your material. This is the single most valuable habit the lesson teaches: before you rely on any tailored assistant, ask what it was given.
One last reflection, away from the keyboard: the next tailored assistant you meet that you did not build, at work, in a support chat, inside your software, has standing instructions you cannot see and documents you did not choose. You now know exactly what to ask about it. In the next lesson, capability turns into judgment: whether AI should do a given task at all.
Flashcards
Section titled “Flashcards”Q. What is a system prompt?
Instructions and context written once, in a special place, and silently attached to every conversation that follows. Whatever you put there stays “valid for every single interaction.”
Q. Why do your instructions normally vanish between conversations?
Instructions and context ordinarily wash away: close the chat, and the machine’s next conversation starts blank. That is why re-pasting the same paragraphs feels like being a receptionist for your own assistant.
Q. What belongs in standing instructions, and what does not?
Only what is stable: who you are, how you like output, the job. Per-task detail belongs in the task’s prompt, because a system prompt applies to every conversation.
Q. How does a system prompt fit the prediction-machine model?
The machine writes one word at a time from everything in front of it. A system prompt simply guarantees certain words are always in front of it, a job description quietly stapled to page one.
Q. What is retrieval augmented generation, RAG for short?
You give the AI documents, websites, or databases that matter to you, a pile called a knowledge base. On every question the system searches it, quietly attaches the most related passages to your prompt, and the machine answers from the combined text.
Q. Why is retrieval called context, industrialized?
Lesson 2 called context the part of the prompt almost everyone skips and the part that helps most. Retrieval fetches the right pages for you, on every question, automatically. Same prediction machine, with an automated librarian standing in front of it.
Q. What are the two benefits of grounding an assistant in your own documents?
Your documents can be newer than the model’s training cutoff, and grounding in real pages can help cut down on made-up answers. Fewer, not zero: a grounded assistant can still misread its ground.
Q. What was the admissions assistant good at, and where did it fail?
Good at pulling answers from its material in a friendly, professional tone. Bad at anything its documents did not cover. A tailored assistant is a specialist; its lane is exactly as wide as what you fed it.
Q. What is fine-tuning, and who does it?
Developers train the model further on examples from their field, so the model itself changes, not just what it reads. It is “typically harder and more expensive,” a job for specialist teams; system prompts and retrieval are the two methods most of us can use.
Q. How should you treat 2024's custom GPTs and Copilot demos?
As explicitly dated historical markers, like the 2022 launch date in lesson 1. Those products kept changing while the pattern spread. What survived: standing instructions, retrieval, and ordinary people tailoring assistants without code.
Q. Why does it matter who tailored an assistant?
Because, as the course puts it, “the designer has a lot of control on how it’s going to respond.” Someone wrote the instructions and chose the documents, and that someone had goals. A store’s assistant is tuned to sell. Every tailored assistant has a tailor; ask who.
Q. What can these systems take in beyond typed text?
As of mid-2026, most major assistants can look at an image you give them, many can hold a spoken conversation, and a spreadsheet question can be asked in plain English. New doors into the same machine.
Q. How does voice mode work 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.
Q. What stayed the same underneath the job description, the filing cabinet, the eyes, and the voice?
In the course’s words, “the logic behind the tools is the same”: a prediction machine writing one word at a time from an enormous body of human writing, tuned to be helpful. A tailored assistant is not a smarter machine. It is the same machine, better briefed.