Skip to content
NTPU AI Learn

AI 102 · Module 1

From Prompt to Context

Why can a well-written prompt still lead to a wrong answer?

Start Section 1
  • Duration50 min
  • LevelFoundation
  • PrerequisiteAI 101 (Claude AI Workforce Fundamentals) recommended
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Explain the difference between a prompt and context
  • Understand that an AI answer depends on the instruction plus the context
  • Recognize errors caused by missing information, and tell prompt problems, context problems and situations that need human judgment apart
  • State the basic purpose of Context Engineering

The Problem

8 min

A Good Prompt, a Wrong Answer

In AI 101 you learned to state a task clearly. Yet many people hit this situation: the instruction is clear, the format is spelled out, and the answer is still unusable. Making the prompt longer rarely helps.

Work scenario

Please analyze the main risks in this contract.

Nothing is wrong with that sentence. But to analyze well, the AI also needs to know:

  • The full contract, not just a few paragraphs
  • How your organization judges risk
  • The applicable law and earlier versions
  • Background on this particular client

Warning

When that information is missing, the AI does not always say "I do not know". It may infer from typical cases and hand back an analysis that reads well. Reading well is not the same as being right.

Key takeaway: A clear instruction does not mean the AI has the information it needs to do the task.

Core Concepts

8 min

Prompt versus Context

A prompt mainly describes how you want the AI to work. Context (上下文) is the information the AI can use while doing the task. They do different jobs, and you need both.

Prompt (提示詞)

  • What to do: the task
  • How to do it: steps and method
  • Output format
  • Constraints

Context (上下文)

  • Materials such as the contract or meeting notes
  • Organizational policy and risk standards
  • Earlier versions and client background
  • Where the work currently stands
The basic model: Task + Instruction + Context = AI Output
  1. Task: the work to be done this time
  2. Instruction: how you want the AI to do it and in what format
  3. Context: the information the AI can rely on
  4. AI Output: shaped by all three, then reviewed by a person

So when an answer is disappointing, ask not only "how should I change the prompt?" but also "does the AI have the information it needs?"

Check yourself

A team complains that AI risk analyses of contracts are vague. The prompt already states the task, format and limits, but you notice only the cover page and table of contents were pasted in. What should you do first?

A team complains that AI risk analyses of contracts are vague. The prompt already states the task, format and limits, but you notice only the cover page and table of contents were pasted in. What should you do first?

Key takeaway: The prompt decides how to work; the context decides what to work from. The output depends on both.

7 min

What Context Engineering Does

Context Engineering (上下文工程) is the deliberate decision of what information an AI should see for a task, where it comes from, and when to provide it. It does not mean pushing all your data at the AI.

Done

Good context is not more context. It is the right information, trusted, sufficient and safe, provided at the right time.
  1. What information does this task really need?
    Think about the task first, then the data, not the other way round.
  2. Where does it come from, and can it be trusted?
    Is there an official source, and is it the current version?
  3. Is it appropriate to give to the AI?
    When personal data, confidential material or unauthorized documents are involved, stop and decide first.

These three questions run through the whole course. AI can help organize and analyze, but choosing what to provide, judging whether a source is trustworthy and deciding whether to adopt the result remain your responsibility.

Key takeaway: Context Engineering is judging what information is worth letting the AI know, not stuffing in more.

Demonstration

10 min

Demonstration: One Prompt, Three Contexts

The same administrator asks the AI the same question, but each time the context is different. All data below is illustrative and for demonstration only.

The fixed prompt

Please judge whether this purchase complies with policy: 8 office laptops, total NT$240,000, only one quotation obtained.

Possible results under three contexts (illustrative)
CaseContext providedLikely AI answerHow far to trust it
ANo policy document at allGuesses it "should comply" and notes that purchases usually need approvalNo basis; it is a guess
BA two-year-old policy (quotes required above NT$30,000, but not three)Concludes one quotation is enough and only department approval is neededHas a basis, but the basis is out of date
CThe current official policy (three quotes required above NT$50,000)Points out it does not comply; three quotes are needed before reviewHas a basis with a clear source and version

Discuss:

  • Did the prompt change?
  • How did the context change the answer?
  • Which answer deserves the most trust, and why?
  • If you only saw the text of answer A or B, could you tell something was wrong?

Info

All three answers read fluently and sound confident. What differs is the information they rest on. A user needs to know what the AI relied on, not only what it said.
Check yourself

With no policy document provided, the AI confidently says "this purchase complies with policy". How should you treat that answer?

With no policy document provided, the AI confidently says "this purchase complies with policy". How should you treat that answer?

Key takeaway: With the same prompt, a different context can give a completely different answer. Whether to trust it depends on what it rests on.

Practice

10 min

Practice: Prompt, Context or Human Judgment?

Below are 8 illustrative cases. For each, choose the main cause as you see it, then discuss with a peer or the instructor. Some cases can reasonably be read more than one way; what matters is that you can explain your reasoning.

8 cases (decide for yourself)
CaseWhat happenedThink about
1The instruction says only "tidy this up", and the AI returns three pages of running notesDid the user say what they wanted?
2The instruction clearly asks to compare two suppliers, but only one supplier's quotation was pasted inDid the AI receive what a comparison needs?
3The AI cites a travel policy that has been retired; the instruction itself was fineWhich version was it, and who keeps it updated?
4The instruction asks to "list the risks in a table" and the data is complete, yet the AI writes long proseWas the instruction followed clearly?
5No data is attached and the instruction is vague; the AI answer is off targetDo both need strengthening?
6The AI gives a sound recommendation from the current policy, but this case is about whether to make an exception for a long-term vendorCan AI make that decision for a person?
7A 40-page attachment is pasted in for one small question, and the AI misses a key clause on page 27Is there too much material, or too noisy?
8The AI summary says "the client agreed", but the original email only said "we will look into it"Can the summary be trusted, and who checks the original?
  • Prompt Failure: the instruction was unclear or not followed
  • Context Failure: the AI lacked, received wrong, or received too much information
  • Both: problems on both sides
  • Human Judgment Required: information and instruction are fine, but a person must decide or take responsibility
Check yourself

The AI correctly summarizes the rules for a purchase from the current official policy. But this is an urgent purchase from a long-term vendor, and whether to grant an exception is not covered by the policy text. Which situation is this closest to?

The AI correctly summarizes the rules for a purchase from the current official policy. But this is an urgent purchase from a long-term vendor, and whether to grant an exception is not covered by the policy text. Which situation is this closest to?

Key takeaway: When something goes wrong, first tell apart: the instruction, the information, or a decision that needs a person.

Recap

7 min

Recap: From "How to Ask" to "What to Know"

Done

What is this AI task really missing: the instruction, or the right context?

AI 101: How do I tell the AI?

  • Define the task
  • Write a clear instruction
  • Review the output

AI 102: What should the AI know?

  • Decide what information is needed
  • Judge whether a source is trusted and current
  • Decide what should not be provided

The next module takes context apart: what kinds of information might an AI need to complete a task?

Check yourself

A colleague says: "If I upload every related file to the AI, context problems go away." Which reply best fits this module?

A colleague says: "If I upload every related file to the AI, context problems go away." Which reply best fits this module?

Key takeaway: Good context is the right information, trusted, sufficient and safe, provided at the right time.

Save as PDF

What next?