AI Context Document: What Your Marketing Assistant Needs to Know

An AI context document gives your assistant the business information it needs to do a particular kind of work. For marketing, that includes what you sell, who it is for, what you can prove and what you should not claim.

The useful version is specific enough to inspect. Someone should be able to read a statement and tell whether it is current, where it came from and who can correct it.

A long company description is not necessarily useful context. It may contain ambitions presented as facts, an audience definition nobody has tested, and an offer that changed months ago.

Start with the information needed for one recurring task. Keep confirmed facts separate from assumptions. Add a clear route to the source when a detail needs checking.

This guide uses Definitely Human’s own separation of business facts, customer hypotheses and voice rules. The fictional examples are labeled and demonstrate document structure, not customer outcomes.

What belongs in an AI context document?

For a marketing assistant, begin with the business, offer, audience, proof, writing rules and boundaries. Each section should answer a practical question the assistant would otherwise have to guess.

Include enough detail to support a decision. “We help businesses grow” does not tell a writer which problem to explain or which service to describe. A plain account of what the service produces is more useful.

For example, Definitely Human builds AI marketing systems around a business’s knowledge, tools and recurring work. That description is different from saying we build the underlying model or sell an advertising service. Those distinctions affect the claims a marketing assistant should make.

The document does not need to carry every detail of every service. It needs to point to the right reference and explain when that reference is relevant.

How do you separate facts from assumptions?

Write the status beside the statement. Do not expect a future reader or assistant to infer which parts of a polished document are well supported.

A confirmed business fact can come from the current offer specification or the person responsible for delivery. An audience assumption might come from a small number of conversations and still need testing.

For a fictional consulting firm, the distinction could look like this:

The same idea applies to your own results. If a number belongs to a specific engagement, preserve its scope. Do not turn a result from one situation into a guarantee for every future customer.

At Definitely Human, the business reference and customer reference have different jobs. The first records decisions about the company. The second can hold hypotheses about buyers. Keeping those roles separate helps avoid writing an internal belief as though a customer confirmed it.

An AI context document should preserve uncertainty where uncertainty exists. The aim is to make the work accurate enough to review, not to make every field appear complete.

Use “unknown” when information is missing. A visible gap is actionable: someone can verify it, remove the claim or narrow the task. A plausible invented answer is harder to spot.

Which sources should you use first?

Start with current sources that directly support the task: the approved offer, a delivery checklist, published writing, permission-cleared examples and recent decisions from the responsible owner.

A website is useful input, but it is not automatically the current authority on every detail. An old landing page may still advertise a retired service. A deck can contain a price that was accurate when it was presented.

For each important statement, record the source and when someone last checked it. If the source is a live document, include its location. If the source is a decision, record the decision and its owner rather than relying on someone’s memory of a meeting.

Source quality also depends on what you are trying to establish. A sent message can show how someone writes. It may not prove that a pricing promise in that message remains available today.

We learned a related lesson while building our voice reference: a prepared script and a delivered transcript answer different questions. The transcript is evidence of the delivered words. The unused portions of a script cannot establish what was said.

Apply that discipline to business context. Label a proposal as a proposal, a current policy as a policy, and a result as a result with a defined scope.

How much detail should the document include?

Include the detail that changes the work. Leave out material that does not help the assistant make the decisions required for the task.

For an article about preparing for a service assessment, the writer may need the checklist, common misunderstandings and next step. The company’s complete meeting archive adds little unless a particular decision is missing from the current reference.

Begin with one readable AI context document. Split it when different parts have different owners, change at different rates, or are relevant to different tasks.

A useful structure might be:

• An entry page describing the work and pointing to current references.
• A business reference containing the offer and boundaries.
• An audience reference containing observations and hypotheses.
• A voice reference containing examples and writing rules.
• A short task brief containing today’s assignment.

This is a proposed structure, not a requirement to create five files immediately. If one page remains easy to maintain, keep it as one page. The reason to split is a practical problem, not a desire to make the workspace look sophisticated.

Avoid copying the same price, promise or policy into several files. Copies drift. Link to the owner of the fact and include only the excerpt needed for the current task.

What does a useful business entry look like?

A useful entry gives the assistant a fact, its scope and any action required before using it. It does not merely provide a slogan to paraphrase.

Consider this fictional service description:

We offer a migration assessment. It reviews the current site, its integrations and the questions that must be resolved before a delivery plan can be prepared. It does not itself commit the team to a migration date. The service owner approves scope and timing after the review.

This gives a writer enough information to explain the offer without inventing a guarantee. It also names the decision that remains outside the assistant’s role.

Now compare a weaker entry:

We provide seamless, innovative digital transformation for ambitious companies.

The second gives the writer adjectives without defining the service. It may encourage a fluent description, but it does not constrain what the assistant says the company will deliver.

An entry can be short and still carry the necessary distinctions:

Use this level of precision for the facts that regularly appear in public work. You do not need an elaborate record for every minor wording preference.

How should the task brief connect to the reference?

The standing reference explains the business. The task brief explains what you want now.

For example, the reference may describe an assessment service and its boundaries. Today’s brief might ask for a guide aimed at operations leads who are preparing their first information packet.

Keep the desired format, topic, deadline and campaign-specific destination in the brief. Those details may change tomorrow while the underlying service stays the same.

An example task specification could say:

Draft an educational explanation of what a buyer should prepare before an assessment. Use the current service checklist as the source. Write for someone who has not met us. Identify unanswered questions before drafting. Do not infer delivery dates or guarantees.

This is an illustrative assignment, not a complete downloadable workflow. Adapt it to the actual source and review rules for your business.

Specify how the assistant should handle disagreement. If the task brief names an old offer while the reference describes the new one, the assistant should report the conflict for resolution. Quietly choosing whichever version sounds more convenient hides a business decision.

The AI marketing workflow guide shows where this input check belongs before drafting and how to retain corrections after review.

How do you test whether the context is useful?

Test the document against the kind of task it is meant to support. Ask a question with an answer you can verify, then inspect whether the assistant uses the relevant current source and respects its limits.

For the fictional assessment service, test a straightforward question, a missing-information question and a request that would violate a claim boundary.

These tests show whether the context supports the relevant decisions. They do not establish that the assistant will always behave correctly. Keep reviewing consequential work.

If an answer fails, investigate the cause before adding more text. Was the document unavailable? Was the source ambiguous? Did an old instruction contradict it? Did the task ask for something the reference does not cover?

Adding a paragraph is appropriate when a fact is missing. It is less useful when the real problem is that the assistant never received the file.

Ask the assistant to identify the source it used, then verify that source yourself. A confident statement that it has read the reference is not the same as a correct answer grounded in it.

What should stay out of the shared reference?

Keep information out when it is unnecessary for the work or inappropriate for the tool and people who will receive it. A recurring marketing reference rarely needs account credentials, payment details or an entire private customer record.

Use a permission-cleared summary or a redacted example when the writing lesson can be preserved without identifying the person. Keep separate client material in its own workspace rather than mixing it into a reusable reference for another business.

Also keep unresolved internal debate distinct from current instructions. You may want to preserve why a decision changed, but the current position should be obvious. Otherwise the assistant may combine mutually exclusive versions into a confident new answer.

Do not put every temporary promotion into the standing document. A short-lived campaign belongs in the assignment or a dated campaign reference. If it must appear in the shared context, identify when it stops applying and who will remove or supersede it.

The goal is appropriate context, not maximum context. Every additional claim creates something that may need checking later.

How do you maintain the document without creating another chore?

Update it when a relevant decision changes. An offer change, a retired claim, a new delivery boundary or a repeated correction should trigger a review of the affected section.

A scheduled review can catch things you missed, but it should not be the only mechanism. Waiting for a quarterly cleanup leaves a known wrong statement in circulation.

Give each important reference an owner. That owner does not have to make every edit, but someone must be able to resolve disagreements about what is current.

Preserve the reason for meaningful changes. A short note that says “replaced the old scope after the service changed” is more useful than a date with no explanation. Keep lengthy history outside the working instructions so the assistant does not have to reconstruct the current policy from a timeline.

When you correct output, decide where the correction belongs. A fact correction belongs with the fact. A style correction belongs in the AI brand voice reference. A change that applies only to one reader belongs in the task brief.

What should your first version accomplish?

Your first AI context document should let a competent reviewer answer three questions: what information the assistant received, what remains uncertain, and which decisions still belong to a person.

Build it for a task you already repeat. Verify the offer description, collect the relevant evidence, label the audience assumptions and point to the writing rules. Leave fields unresolved when you cannot support them.

Then use it. A reference that never reaches a real task cannot show you which parts are useful or missing.

If you are starting with content, choose a customer question from the Content Idea Engine and run it through the AI marketing workflow. Keep the factual reference beside the draft during review.

The document has done its job when the work is easier to explain and correct. Its value comes from the decisions it supports, not from the number of pages it contains.