Give It Context

The model cannot use what it cannot see

Three modules, three questions:

1. Module 1 — Who decides what happens next?
2. Module 2 — Which decisions should the human make before the work begins?
3. Module 3 — What information should be available when the remaining decisions are made?

Module 2 gave the agent a bounded job. But the job alone does not contain all of the information needed to perform it. For the fictional briefing, the agent might also need current product information, documentation, information about limitations, and information about the intended audience.

The information available while the model makes a decision is its context. But “available” should not mean “everything we can find.”

Three more words to learn

These are the beginner versions. Anthropic’s engineering guidance defines context far more precisely, and in units this course does not need. The Sources section gives the technical definition beside ours.

Same job. Different context.

The job does not change in this module. Only the information supplied to perform it does.

What follows are three descriptions of what information would be made available. They are not model transcripts, and nothing here was run through a model.

Context A — Almost nothing

Context B — The whole folder

Context C — Selected for the task

The job and the context do different work

A good job does not make weak context strong. Good context does not make an undefined job clear.

Give it something to consult

Not everything you supply is an instruction. The official product documentation is not telling the system what to do; it is information the system can look things up in while carrying out the instructions it already has.

Mixing the two up is a common way for a task to go sideways — a document pasted in without explanation can read as a new set of orders rather than as something to check against.

Help the agent tell one source from another

Useful context keeps enough identifying information for its own reliability to be judged: the source, the title, the date where one is available, and the relevant section or where the material came from.

This matters because a folder flattens things. In Context B, an undated marketing PDF and the current official documentation arrive looking equally authoritative — two files, side by side, with nothing to separate them. An undated brochure should not silently outrank current documentation just because both were in the same folder.

Keeping provenance visible preserves the information someone needs in order to weigh one source against another, which is a separate concern from citation formatting.

Context has limits

A model does not have an unlimited pile of information continuously in view. The working space is finite, which means something eventually has to be left out — and if you do not decide what, something else will.

Products manage this differently. Longer-running systems may:

• select relevant information rather than carrying all of it;
• summarize prior work;
• retrieve material only when it is needed; or
• drop information that is no longer useful.

Not every agent does each of these, and how any particular product handles it is an implementation question this course does not need to answer. This is another reason “give it everything” is not a durable strategy.

There is no number to memorize here. Published limits differ by model and change often, and calculating them is not the skill this module is teaching. The useful idea is simply that the working context is finite, so selection matters.

Context does not stay frozen

So far this module has described context as though it were assembled once, before the work starts. For a single response that is roughly true. For agent work it is not.

As an agent works, new information becomes available:

1. it begins with the job and the reference material;
2. it examines one source;
3. the result becomes relevant to its next decision;
4. it may decide another source is needed; and
5. those intermediate results become part of what shapes later choices.

Agent work is not only “prompt in, answer out.” The information available for the next decision can change as the work progresses.

A later module returns to what that means in practice.

Missing context should stay missing

Suppose the briefing needs current information about how the platform handles uploaded documents. None of the supplied reference material addresses it, and no permitted tool can retrieve it.

The appropriate behavior is to identify the missing information — not to fill in the most plausible answer.

This is the Module 2 stop condition doing its job. A gap in the context is a fact about the task, and reporting it is a real result. The most plausible answer, written confidently, is worse than no answer, because it looks exactly like a researched one. Knowing what is missing is part of using context well.

Build the context set

Your turn. You are preparing the same one-page faculty briefing. These eight items are available:

1. The faculty member’s request and desired one-page format
2. The platform’s current official feature documentation
3. The platform’s current official privacy and data-use page
4. The current pricing and access page
5. A positive independent review published two years ago
6. An 80-page report about AI regulation that never discusses this platform
7. A private email thread about an unrelated technology
8. An undated product brochure found in a shared folder

Sort each one into a category before reading on. Two of the eight are genuinely arguable, so resist the urge to make this a tidy list of right files.

Three categories

Include

This belongs in the working context for this job.

Leave out

This should not be part of the task.

Investigate first

This might belong, but something needs checking before it can be relied on.

One possible review

Include

• 1. Faculty request. It defines the actual job and audience.
• 2. Current official feature documentation. Directly relevant primary information.
• 3. Current official privacy and data-use page. Directly relevant to understanding important platform behavior.
• 4. Current pricing and access page. Include if access or cost matters to the decision; otherwise it may not deserve working-context space.

Leave out

• 6. Unrelated AI-regulation report. Length does not make it relevant.
• 7. Private unrelated email. It is unnecessary for this task and should not be exposed merely because it is accessible.

Investigate first

• 5. Two-year-old independent review. It may provide useful perspective, but check whether its claims are still current, and decide whether secondary material belongs in a job built around primary sources.
• 8. Undated product brochure. Determine source and date before relying on it.

Reasonable people may make different choices about items 4 and 5. The important skill is being able to explain why the information belongs in the task.

Relevance beats accumulation

It would be easy to read this module as “less context is better.” That is just as wrong as “more is better,” and Context A is the proof: a starved task fails as surely as a flooded one. The goal is enough context to support the decisions the agent needs to make — without unnecessarily adding information that is irrelevant, stale, confusing, or inappropriate to expose.

Neither “more” nor “less” is the variable that matters. Fit is.

The job tells you what context matters

Context selection is not a separate skill from Module 2. It begins with the job, which is why the job comes first.

• If the job requires three primary sources, then a folder full of commentary does not satisfy the requirement, however thick it is.
• If the job forbids uploading Stanford materials, then those materials do not become appropriate context merely because they might improve the answer.
• If the stop condition says identify unsupported gaps, then missing information does not have to be hidden — the job has already said what to do about it.

The job helps determine what belongs in context. Context helps the agent pursue the job.

Context is more than documents

One last widening, because it is easy to finish a module like this thinking context means uploaded files.

Depending on the system, context may include:

• the user’s current request;
• instructions;
• reference materials;
• prior conversation;
• results returned by tools; and
• other information the host makes available.

Notice that two of those are not things you upload at all. The results a tool returns become context the moment they come back, which is the same point the earlier section made about intermediate results.

“Prior conversation” here means the current piece of work. This module is about the information available for the decision being made now. Whether a product carries anything over between separate sessions is a different question, with different controls, and this course does not assume any product does.

What to notice

Context is the information you make available for the decisions the agent needs to make, which is narrower than everything you could hand it.

• Too little — important decisions lack evidence.
• Too much — relevant information competes with material that may be stale, irrelevant, confusing, or inappropriate.
• Fit for the job — enough useful information is available while unnecessary material stays out.

Before you continue

You are ready for Module 4 if you can:

• explain context without equating it with a prompt;
• distinguish instructions from reference material;
• identify missing context;
• identify information that does not belong merely because it is available;
• explain why source and date information matters;
• explain why context is finite; and
• choose a focused set of information for a defined job.