PrerequisiteAI 101 recommended, plus Modules 1 to 3 of this course
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Explain why not every piece of information belongs directly in the prompt
Choose Include, Retrieve, Remember or Exclude for a piece of information and justify the choice
Describe retrieval and Retrieval-Augmented Generation (RAG) in one sentence, and what still needs a human
Build a simple Context Decision Table for a work scenario
The problem
10 min
Is pasting everything in the best approach?
The first three modules taught you to judge which information deserves to reach the AI. This module asks the next question: once you have chosen it, how should it be delivered?
Core question
Should all context go straight into the prompt? The answer is no.
Suppose you ask an AI to help assess a purchase. You have today's request, a whole folder of procurement policy, your manager's usual report format, and the login password to a supplier portal. These four items do not all get the same treatment.
Four context strategies
Include: needed for this task, short and stable
Retrieve: large, frequently updated, and only a small part is needed each time
Remember: likely to stay useful in the future
Exclude: exists, but should not be given to the AI
Info
Good context is not more context. It is the right information, at the right time: trustworthy, sufficient and safe.
Key takeaway: After choosing what to provide, decide how to provide it.
Four context strategies
10 min
Include: put it in for this task
Include is the most direct option: place the information itself into this conversation. It suits information that meets all of the following.
Short: it will not crowd out other important content
Highly relevant: it directly affects this result
Stable: it will not change during this task
Necessary: the task cannot be done well without it
Good candidates for Include
The transcript of this meeting, the contract clause under review, the work instruction for this task.
A quick test: if you can confirm within two minutes that "this content bears directly on this task", it is usually a candidate for Include. If you would have to dig around to know whether it is relevant, consider Retrieve first.
Check yourself
A company has an 800-page procurement policy library, updated monthly, and each assessment usually needs only a few clauses. Which strategy fits best?
Key takeaway: Include suits information that is short, highly relevant, stable and necessary for this task.
10 min
Retrieve: fetch it when needed, and the idea of RAG
Retrieve (retrieval, 檢索) means pulling a small relevant part out of a large body of information only when it is needed. It suits material that is large, changes often, or is only partly needed each time.
A basic retrieval mental model
Question: the user asks what they need answered
Retrieve: find the relevant content in the knowledge base
Add to context: place what was found into this context
Model: answer based on the question and the retrieved content
This pattern is often called Retrieval-Augmented Generation (RAG, 檢索增強生成). In this course you only need the idea: look up material first, then answer. How to build an index or a system is outside this course.
Warning
RAG only guarantees that a lookup happened. It does not guarantee the right item was found, that nothing is missing, or that it is the latest version. Retrieved content is still context: check it for relevance, authority, freshness and sensitivity, and have a person verify the source.
RAG is also not the whole of Context Engineering. It is one way, among four strategies, to handle large bodies of information.
Key takeaway: Retrieve means finding relevant material before answering. Finding something is not the same as finding the right thing, so people still verify the source.
10 min
Remember and Exclude: what to keep, and what not to give
Remember
Information likely to stay useful in future work
Helpful across sessions
For example: user preferences, project names, work habits, usual formats
Memory does not mean keeping every conversation forever
Exclude
Information that exists but should not be given to the AI
For example: passwords and API secrets, unrelated personal data, unauthorized documents, unnecessary client data
The test is whether this task needs to know it
Relevant does not mean allowed: check sensitivity and authorization first
Warning
Do not paste company secrets, client personal data or unauthorized information into an AI tool. If you are unsure whether you may share something, the default is Exclude, and you ask the owner.
Memory needs care too: an old preference may have changed, and a remembered item may simply be wrong. Reviewing what is remembered is the user's responsibility.
Check yourself
Your manager wants every report to open with the conclusion, followed by three key points. Which strategy fits best?
Key takeaway: Remember is for information that stays valuable across tasks. Exclude is for information that should not be given even though it exists.
Decision table and practice
10 min
The Context Decision Table and practice
Put the four strategies side by side and you have a Context Decision Table. The table below uses an illustrative procurement scenario.
Context Decision Table (illustrative)
Information type
Include
Retrieve
Remember
Exclude
Instruction for this task
✓
Large policy library
✓
User's format preference
✓
System login password
✓
Transcript of this meeting
✓
Frequently updated product price
✓
Info
A table is only a starting point. The same kind of information can get a different answer in a different task, so write a reason for every row.
1List the information Write down 6 to 8 pieces of information your work scenario needs.
2Choose a strategy Pick one of Include, Retrieve, Remember or Exclude for each.
3Write the reason One sentence each, for example "updated often, so Retrieve".
4Check sensitivity Make sure nothing that should not be shared ended up under Include or Remember.
To finish a price comparison, the AI would need the login password to a supplier portal. What is the best handling?
Key takeaway: A decision table puts the four strategies and your reasons side by side so you can defend your choices.
Recap
10 min
Recap
Done
Not everything belongs in the prompt. Use size, how often it changes, repeat value and sensitivity to choose Include, Retrieve, Remember or Exclude.
The four strategies at a glance
Strategy
Suits
Watch out for
Include
Short, highly relevant, stable, needed now
Too much causes overload
Retrieve
Large, frequently updated, partly needed each time
May find the wrong item, miss items, or fetch an old version
Remember
Preferences and background that stay valuable across tasks
Needs periodic review; content can go stale
Exclude
Passwords, unrelated personal data, unauthorized documents
Relevant does not mean allowed
Choosing a strategy is judgment, not a formula. A person decides what the AI receives and stays responsible for the result. The next module looks at how AI fails when context is designed badly.
