Today is not about writing better prompts.
It is about helping AI know what you want before you ever hit send.
You already know prompt engineering — the seven elements from Lesson 2. This is the other half, and it is the half almost nobody learns.
"Help me plan my college experience."
A perfectly clear instruction. The AI understands the task completely — and still knows almost nothing about you.
Nothing yet. The AI is answering that one sentence with no idea who is asking it.
A language model answers by predicting the most probable useful response. With no information about you, the most probable useful response is the one that would be least wrong for the largest number of people. That is the definition of generic. It is not a flaw in the model — it is the model correctly hedging against everything it does not know.
Give it real detail and the probable answer changes, because the population it is averaging over shrinks to something much closer to one person.
A context tool is an AI chat whose job is not to answer your question but to gather the background it will need to answer it well later. Today, we will use a context tool that interviews you and turns your answers into a file you keep. The concept is to allow AI to have a better understanding of you. A useful interview does not simply record your answers — it challenges the vague ones. Watch for this pattern today,because it is the difference between a document that describes you and one that just flatters you.
"I want to be successful."
"What would success actually look like to you?"
"I want a good career."
"Good how — interesting work, income, creativity, stability, helping people, independence?"
What you build today is yours. You will not submit it. You may share it if you want feedback, and that is entirely optional. Skip any question you don't want to answer. Nothing here requires you to discuss trauma, medical history, family problems, finances, or relationships — and you should not put anything into an AI system that you would be uncomfortable having stored on someone else's server.
Before we get started, you need to understand the Markdown file. You have been reading it for months without knowing its name. Those stray asterisks that show up when you paste ChatGPT output into a Google Doc? That is Markdown, unrendered.
Markdown is plain text with a handful of punctuation conventions that mean "this is a heading," "this is bold,"
"this is a list." That is the entire idea. A Markdown file is a .md
file, and if you strip away every tool that renders it, you can still read it. It is just typing.
Compare that to a .docx, which is a zipped bundle of information that only opens correctly in software that agrees to understand it. Markdown has no version number, no subscription, and no company that can discontinue it.
##
headings is easy. Headings are not decoration — they are load-bearing.Everything you need for professional work is in this box. Edit it. Break it. See what each mark does.
| You type | You get |
|---|---|
| # Heading | The biggest heading. One per file — it's the title. |
| ## Section | A section heading. This is the one that does the real work in AI prompts. |
| **bold** | bold — for emphasis and labels |
| *italic* | italic — for titles, asides, uncertainty |
| - item | A bulleted list. One dash, one space, one item per line. |
| 1. item | A numbered list. Use it when order matters. |
| > quote | A block quote — perfect for source quotes in reporting notes |
| [text](url) | A link. Text in square brackets, address in parentheses. |
Two more worth knowing: `backticks`
mark something as literal text, and ---
on its own line draws a horizontal rule.
VS Code is the free text editor we downloaded for this class. Today you will use it to make real files in a real folder, because a knowledge base that lives inside a chat window is not a knowledge base. It is a conversation you are about to lose.
Here is what you are building. You will not finish all of it today, but you will start it.
Comm-101/ # you already made this └── ai-workbench/ # make this today ├── context/ # who is asking │ ├── my-college-roadmap.md # today's lab │ ├── my-college-context.md # today — paste this one │ └── my-beat.md # later ├── prompts/ # what to do │ ├── interview-me.md # today │ └── show-notes.md # later ├── sources/ # what is true └── README.md # what's in here and why
Those three folders are not tidiness for its own sake. Each one holds a different kind of thing you hand to an AI, and each answers a different question the AI cannot answer on its own.
Files about you. Your major, your goals, what you have already tried. You paste one in at the start of a conversation so the AI stops answering as if you were the average student.
Instructions that worked well enough to use again. Writing a strong prompt takes real effort, so you save it here once and paste it next time instead of rebuilding it from memory.
Raw material you have verified. Interview transcripts, your notes, quotes with a page number. When an answer has to rest on facts rather than the model's guesses, it comes from here.
Notice how that maps onto the last two lessons. Lesson 2 taught you to write what
goes in prompts/.
Today is about what goes in context/.
You will fill sources/
in later lessons, and it is the folder that keeps
an AI from inventing a source on you.
Work through these in order. Each one takes under a minute.
ai-workbench.
Type the name exactly as
shown: all lowercase, one hyphen, no spaces. Every instruction for the rest of this lesson assumes that
spelling.ai-workbench:
context,
prompts,
and sources.
Same rule — lowercase, no spaces.
ai-workbench
itself, and click Open. Your three folders should now be listed in the sidebar
down the left side of the window.
README.md
and save it directly inside ai-workbench,
not inside one of the three folders. The .md
on the end is what tells VS Code the file is
Markdown.# My AI Workbench
— and save with Cmd S on a Mac or Ctrl S on Windows.Markdown looks like stray punctuation until something turns it into formatting.
With README.md
open, press Cmd K then
V on a Mac, or Ctrl K then V on Windows. A preview opens beside your text, and the line you typed should appear
large and bold. To see only the preview with no editor next to it, use Cmd Shift V or Ctrl Shift V.
Your sidebar should show ai-workbench
at the top, the three
folders beneath it, and README.md
below those. If the sidebar looks empty, you opened a file
instead of a folder — go back to step 4. If your heading still shows the #
in the preview,
you are missing the space between the #
and the first word.
context/
is something you paste into an AI, not something you read. You are not writing notes for yourself. You
are writing a briefing document for a very capable assistant with no memory.
You are going to run a controlled experiment on yourself. Same question, asked twice, forty minutes apart. The only thing that changes in between is what the AI knows about you.
Open a brand new conversation in Claude, ChatGPT, or Gemini. Do not explain your major, your interests, your goals, or your background. Do not warm it up. Paste exactly this and nothing else:
Read the answer with these questions in mind. Be harsh — you are evaluating the answer, not the AI.
You are about to hand the AI a 2,000-word prompt. Open it and read it first. You are not studying it. You are just seeing that someone built it on purpose, and that every paragraph in it is doing a job.
Download this file:
Save it into your ai-workbench/prompts/ folder, keeping the
name interview-me.md. Then open it in VS Code and
press Cmd/Ctrl Shift V to read it rendered.
What this prompt is doing. Eight techniques are built into it. You do not have to memorize them. You do have to notice that they are there, because you will reuse them every time you write a prompt that matters.
| Technique | How it shows up in the file |
|---|---|
| Give it a job, not a personality | It says to interview you and produce a document. It never says "you are a career counselor." A job title would arrive carrying assumptions about what a nineteen-year-old ought to want. |
| Make uncertainty a legal answer | It lists "I don't know," "I'm not sure," and "Pass" as acceptable up front, and tells the AI to treat uncertainty as useful information instead of a problem to fix. |
| Control the pace | One question at a time, roughly ten to twelve of them, finished in ten to fifteen minutes. Without this the AI hands you twenty questions in one block and you answer none of them well. |
| Show the behavior, don't just name it | Instead of "challenge vague answers," it works an example: if I say I want to be successful, ask me what success means to me. Showing beats describing. |
| Say what not to do | Do not simply agree with everything. Do not diagnose me. Do not invent strengths to be encouraging. Naming the failure you fear is often more useful than naming the goal. |
| Build in consent | Anything sensitive has to be labeled optional, and the AI is told not to press when you skip it. The prompt protects you from the tool. |
| Specify the output | Thirteen numbered sections, each with its own instruction. The AI does not get to decide what shape the finished document takes. |
| End by writing your context file | The last section asks for a 300 to 500 word summary built to paste into other AI conversations. That is the whole point of today, and the prompt produces it for you. |
Start another new conversation — separate from your baseline. Open
interview-me.md, select all of it, and paste the whole thing in as your first message.
Then answer its questions.
Answer in full sentences. Give examples when it asks for one. Say "I don't know" when you don't. Push back if a question feels wrong. Let it challenge you — that is what you told it to do.
Don't perform. Don't write what sounds impressive. Don't invent an interest you don't have — you would be engineering context for a student who isn't you, and the advice at the end would be for them.
When the interview ends, the AI writes My College Roadmap — thirteen sections, from "Where I Am Now" through "Context for Future AI Conversations."
Read it. It should tell you something you did not say out loud. Section 12, What I May Be Underestimating, is usually where that happens.
ai-workbench/context/my-college-roadmap.md
and paste the whole thing in. It arrives as Markdown already — the headings, bold, and lists come across
intact. Press Cmd/Ctrl Shift
V and watch it render.
Why this step matters more than it looks: the chat you just had will scroll away, get archived, or get deleted. The file will not. Everything you did for the last twenty minutes only becomes reusable at the moment it becomes a file.
Section 13 is the payload. It is a 300–500 word block written specifically to be pasted into other AI conversations. That is your first knowledge base entry. Copy it into its own file too:
ai-workbench/context/ ├── my-college-roadmap.md # the full document └── my-college-context.md # just Section 13 — paste this one constantly
Stay in the conversation that has your roadmap in it. Paste the question again — word for word, exactly as you asked it at the start.
The quality of what AI gives you depends on the quality of what AI had when you asked. You did not learn a better prompt today. You learned that the prompt was never the bottleneck. In four weeks you will do this again for a podcast, and again for a client brief, and the move will be identical: build the context first, save it as a file, paste it in.
You just spent forty minutes giving an AI context. Here is the uncomfortable part: it is going to lose it.
In Lesson 1 you learned that every model has a context window — a hard ceiling on how many tokens it can hold at once, and a whiteboard that drops old writing off the left edge when it fills. That was the concept. This is the consequence.
Everything in a conversation shares that window: your instructions, the AI's replies, pasted documents, all of it. Your roadmap is not stored anywhere special. It is just early text — which makes it the first thing to fall off.
Add turns and watch what happens to the instructions you gave at the beginning.
Simplified classroom demonstration. Real windows are far larger and models compress rather than truncate cleanly — but the failure this produces is exactly what you will experience.
It will not announce it. It will just quietly get worse. The signs:
Don't run your whole week in a single thread. Start a fresh conversation per task and paste your context file at the top. A short conversation is a conversation that still remembers everything.
When output drifts, don't type "remember what I said earlier" — it can't. Paste the file in again. Ten seconds of pasting beats twenty minutes of re-explaining yourself badly.
Before a long thread dies: "Summarize everything important from this conversation as Markdown I can reuse." Save it. Start fresh with that summary. You've compressed the thread into a file.
Keep your context in a file you control, not in a conversation you don't. That is the entire argument for the folder you built today — and it is why a knowledge base is infrastructure rather than note-taking.
New terms from Lesson 3 — plus the question that comes with keeping files about people.
The practice of deliberately assembling what an AI knows before you ask it to do anything. Where prompt engineering asks "how should I phrase the request," context engineering asks "what information does this system need in order for the request to be answerable well?" In professional work the second question usually matters more, because a perfectly phrased request built on nothing still produces generic output. Context engineering is why a saved file beats a clever sentence.
The maximum amount of text — measured in tokens — that a model can hold in working memory at one time. Everything counts against it: your instructions, the AI's replies, and any document you paste. Introduced in Lesson 1 as a concept; in Lesson 3 it becomes a practical constraint you plan around. The window is not storage. Nothing in a conversation is saved anywhere you can retrieve it later, which is why the file on your own machine is the only durable copy of your context.
The gradual degradation of AI output over a long conversation as early instructions fall out of the context window. The model does not warn you. It simply starts re-asking questions you answered, violating constraints you set, and sliding back toward its generic default voice. Recognizing drift — rather than concluding the AI "got worse" — is the difference between fixing the problem in ten seconds and abandoning a session that was working fine.
An organized collection of files holding the information you repeatedly need an AI to know — who you are, what you make, who your audience is, what standards you work under, and what facts you have already verified. The distinction from note-taking is the audience: notes are written for you to read later, a knowledge base is written to be pasted into a system that knows nothing. That changes how you write it. You spell out what feels obvious, because obvious-to-you is invisible-to-the-model.
A lightweight way of marking up plain text using ordinary punctuation — # for headings, ** for bold, - for list items. Files use the .md extension. Markdown is the native output format of every major AI chat tool, the standard for technical and editorial documentation, and readable with or without software that renders it. For communications students its practical value is portability: the same file moves cleanly between an AI conversation, a text editor, a CMS, and a collaborator's machine.
A file containing only characters — no embedded fonts, styling, tracked changes, or proprietary structure. A .md file is plain text; a .docx is not. Plain text is the most durable digital format in existence: it has no version compatibility, requires no particular software, and will open in any editor on any system decades from now. For journalists especially, this matters — your notes, transcripts, and source material should outlive whatever app you took them in.
Source is what you type — ## Key quotes. Rendered is what a reader sees — a formatted heading. Markdown files are stored as source and displayed as rendered, which is why the same file can be edited in a bare text editor and still look polished in a preview pane. The stray asterisks you have seen when pasting AI output into a Google Doc are Markdown source that never got rendered, because Docs does not speak Markdown by default.
A folder of saved, reusable prompts stored as files rather than retyped each time. A prompt earns a place in the library when you will need it again — a show-notes prompt, a press-release prompt, an interview-me prompt. Professionals maintain these the way a photographer maintains presets. The alternative, which most people do, is rewriting the same prompt from memory dozens of times and getting a slightly different result every time, with no way to tell which version worked.
The exercise was safe by design — you were told to skip anything you did not want stored, and the interview stayed on academics and skills. But the technique you learned does not stay personal. Within a year, some of you will be building context files about sources, clients, and interview subjects, and pasting those in too.
You promise a source confidentiality. Then you paste your interview notes into an AI to help you structure the story. Your source's name, employer, and unpublished allegation are now on a third party's infrastructure. You did not publish anything — but you did disclose it, and you never told them.
A client shares an unannounced product, a layoff timeline, or a crisis plan under NDA. You build a context file so the AI writes on-brand. The NDA almost certainly did not carve out an exception for pasting it into a chatbot.
And one carried forward from the byline question: the roadmap you generated today was written by an AI about you. It reads as authoritative, and some of it will be slightly wrong — a strength you don't actually have, an interest it overweighted. Read it as a draft of yourself, not a diagnosis. The same skepticism you owe an AI-drafted press release, you owe an AI-drafted description of your own future.
Did the roadmap get anything about you wrong? Where did it flatter you? If an AI that interviewed you for fifteen minutes can produce a confident, wrong description of who you are — what does that suggest about the confident descriptions it produces of people it never interviewed at all?
Lesson 3 · Context Engineering — Giving AI Something to Work With