Week 13: Teaching vs Prompting: 6 Differences That Change Your AI Results

Week 13 · Teaching, not prompting

Teaching vs Prompting: 6 Differences That Change Your AI Results

Zafrullah Khan, Ed.D. · about 8 minutes

2026-11-04

Short answer: Prompting tells an AI model what to do in one moment; teaching gives it organized knowledge and feedback it can use again. Prompting is fast but fragile, while teaching takes more setup and gives results you can repeat and check. The six differences below cover where the knowledge lives, how long it lasts, how you check it, how mistakes get fixed, what size model you need, and who can maintain it.

This lesson pairs with What did your model forget? and the rest of the Learn hub.

The 6 differences at a glance

1. Where the Knowledge Lives: In the Prompt or in Organized Sources?

With prompting, the knowledge lives in what you type. If you want the model to know your refund policy, you paste it in, or you hope the model already learned something close during training. With teaching, the knowledge lives in organized sources the model reads from: documents, examples, and definitions that someone has arranged on purpose.

• Prompting: knowledge is whatever fits in the box today, plus whatever the model remembers
• Teaching: knowledge is a maintained set of sources the model draws on every time
• Why it matters: a model answering from memory can mix up details; a model answering from sources has something to stand on

A simple test shows the difference. Ask the same factual question about your own work three times in fresh chats with no pasted context. If the answers vary, the model is guessing from memory. Now give it a short, clean source and ask again. Answers usually get steadier, because the knowledge has a home.

Teaching does not mean writing a giant document. It means deciding what the model must know, writing it down clearly once, and keeping it in one place.

Trade-off: organizing sources takes time up front. For a one-off question, pasting context into a prompt is fine.

2. How Long It Lasts: One Conversation or Every Conversation?

A prompt lasts as long as the conversation. Close the chat, and the careful instructions you wrote are gone unless you saved and re-pasted them. Teaching lasts across conversations, because the knowledge and examples stay in place for the next question, the next user, and the next week.

• Prompting: each session starts from zero
• Teaching: each session starts from what has already been taught
• Why it matters: repeated work should not depend on someone remembering the perfect wording

Many teams discover this the hard way. One person writes a prompt that works beautifully, then shares it in a chat thread. A month later there are five versions, each slightly different, and nobody knows which one produced last quarter’s report. That is prompting at scale: copies drifting apart.

Teaching flips this. You improve the shared sources, and everyone using the model benefits at once. The improvement compounds instead of evaporating.

Trade-off: lasting knowledge can also mean lasting errors. If you teach something wrong, it spreads to every conversation until you fix it, which is why the next difference matters.

3. How You Check Answers: Read and Hope, or Trace to a Source?

With prompting, checking an answer usually means reading it and deciding whether it sounds right. That works when you already know the answer, which is exactly when you need the AI least. With teaching, a well-built model can point to the source behind each answer, so you can confirm it in seconds.

• Prompting: you judge the answer by how confident it sounds
• Teaching: you judge the answer by where it came from
• Why it matters: confident wording is not evidence; a traceable source is

This is what we mean by checkable AI. A checkable answer comes with a trail: which document, which section, which example. If the trail is missing or does not support the answer, you know to be careful. A taught model should also be able to say “I don’t know” when the sources do not cover a question, instead of filling the gap with a guess.

To try it, ask your model a question and then ask, “Which source supports that?” If it cannot answer, or names something that does not exist, you are relying on memory, not teaching.

Trade-off: source-tracing adds a step for the reader. For low-stakes drafts, a quick read may be enough.

4. How Mistakes Get Fixed: Rewrite the Prompt or Correct the Source?

When a prompted model gets something wrong, the usual fix is to rewrite the prompt: add a warning, add an example, add “do not do X.” Over time prompts grow long and brittle, and a fix for one question can quietly break another. When a taught model gets something wrong, you correct the source it used, and the fix applies everywhere that source is used.

• Prompting: fixes pile up as extra instructions
• Teaching: fixes go into the knowledge itself, once
• Why it matters: one correction in the right place beats ten patches in the wrong place

Think of how you would help a new colleague. You would not hand them a longer list of “don’ts” every time they slipped. You would show them the right reference and explain why. Teaching an AI model works the same way: correct the material, add a clear example, and re-check.

Keep a short log of each mistake, the source you corrected, and the date. That log becomes your re-test list, so you can confirm the fix held.

Trade-off: source fixes require knowing which source caused the error. If answers are not traceable (difference 3), you are back to guessing.

5. Model Size Needed: Does Teaching Let a Smaller Model Do the Job?

Prompting tends to lean on large models, because a large model has memorized more and can cover gaps in your instructions. Teaching shifts the load from memory to organized sources, so a smaller model often has enough to answer well on a focused task. That can mean lower cost, faster answers, and the option to run on a laptop or phone.

• Prompting: bigger models hide gaps with broad memory
• Teaching: a focused, well-taught smaller model can match a larger one on narrow work
• Why it matters: size should follow the job, not habit

This is not a promise that small always wins. Smaller models lose breadth: rare topics, long chains of reasoning, and unusual edge cases. The point is that teaching narrows what the model has to remember, which makes a smaller model a realistic option worth testing.

Before you switch, run the checks in What did your model forget?: compare the smaller and larger models on 20 to 50 real questions and read every answer that differs.

Trade-off: a taught smaller model is only as good as its sources. Broad, open-ended work may still need a larger model.

6. Who Can Maintain It: One Prompt Writer or Anyone Who Knows the Subject?

Prompting depends on prompt-writing skill: knowing which phrasing, order, and tricks get the model to behave. That skill tends to sit with one or two people. Teaching depends on subject knowledge: knowing what is true and explaining it clearly. That skill already exists across your team.

• Prompting: maintenance sits with whoever understands the prompt
• Teaching: maintenance sits with whoever understands the subject
• Why it matters: the people who know the answers should be able to fix the AI

A nurse, an accountant, or a teacher may never write a clever prompt, but each can spot a wrong answer and correct the source behind it. Teaching lets them do that directly, without translating their expertise into prompt tricks. It also makes handovers easier: when the prompt writer leaves, the taught knowledge stays.

Start small. Name one owner per topic, give them a simple way to review answers and edit sources, and schedule a monthly re-check.

Trade-off: shared ownership needs light coordination, such as who approves changes, or sources can drift.

When prompting still helps

Teaching is not a replacement for good prompts. Prompting is still the right tool when:

• the task is one-off, exploratory, or creative, such as brainstorming or a first draft
• you need tone, format, or length changes rather than new knowledge
• you are testing whether an AI model can help at all before investing in teaching
• the stakes are low and you will review the output anyway

Most real setups use both: teaching for what must be right every time, prompting for how a specific answer should look.

Frequently asked questions

What is the difference between teaching and prompting an AI model?

Prompting gives an AI model instructions for one conversation. Teaching gives it organized sources, examples, and corrections it can use in every conversation. Prompting is quick and flexible; teaching is slower to set up but makes answers more consistent, easier to check, and easier to fix.

Is prompt engineering still useful?

Yes. Prompting is still the fastest way to shape tone, format, and one-off tasks, and to test whether AI can help at all. It becomes fragile when you rely on it for knowledge that must be right every time. For that, teaching the model with organized sources is more reliable.

Does teaching an AI model mean retraining it?

Not necessarily. Teaching can mean giving a model organized sources to answer from, clear examples, and a routine for correcting mistakes, without changing the model itself. Some teams also fine-tune, but the core idea is that knowledge lives somewhere you can see, check, and fix.

Can a smaller AI model work well if it is taught?

Often, for focused tasks. When a model answers from organized sources instead of memory, it needs less built-in knowledge, so a smaller model can match a larger one on narrow work. It may still struggle with rare topics or long reasoning, so compare it with a larger model on your own questions first.

How do I check if an AI answer came from a reliable source?

Ask the model which source supports the answer, then open that source and confirm it says the same thing. If the model cannot name a source, names one that does not exist, or the source does not support the answer, treat the answer as unverified. Checkable AI makes this trail visible by default.

Keep learning

• Learn hub
• What did your model forget?
• 7 Checks Before You Trust a Smaller AI Model

Want to talk through teaching a model for your own work? Book a conversation.

Zafrullah Khan, Ed.D., writes about checkable AI, teaching rather than prompting, and practical small models.