Prompt Engineering: How to Communicate Effectively with AI

1. A prompt is part of the model's context

A prompt is the input that tells an AI model what you want it to do and provides information that may help it do the task. In a chat product, the prompt may include your latest message, earlier conversation turns that are still included, attached text, and instructions added by the application.

Lesson 4 introduced the next-token process. The model uses the context it receives to estimate what text is likely to come next. Your prompt helps shape that context.

Consider two requests:

Vague request

Help me with a customer.

More useful request

Draft a polite email replying to a customer who asked whether a weekend service fee applies. Use only the fictional FAQ included below. If the FAQ does not state the fee, say that the office needs to confirm it. Do not invent a price. Keep the reply under 100 words.

The second request gives the model a clearer job and a boundary. It reduces the amount of guessing required.

Prompt engineering is the process of making this kind of improvement deliberately. You describe the job, provide the relevant information, set expectations, and check the result.

Clearer does not mean magical

A carefully written prompt can make the desired response easier to produce. It cannot guarantee that every answer is correct. The model may misunderstand an instruction, overlook part of a source, produce a plausible but unsupported detail, or respond differently when the same prompt is used again.

A prompt is guidance supplied to a model, not a replacement for verified information, software permissions, or human judgment.

A useful question before you type

Ask yourself: If I gave this request to a capable new employee who knew nothing about my situation, what would I need to explain?

That question is a practical starting point for better prompts.

2. State the task and define what a good result looks like

Begin with a clear action. Tell the model what you want it to produce.

Useful task verbs include:

• Draft: write a first version of an email, notice, or message.
• Summarize: shorten supplied material while preserving the important points.
• Compare: identify similarities and differences between supplied options.
• Organize: turn rough notes into a checklist, outline, or table.
• Explain: teach a topic at a stated level.
• Review: check a draft against criteria you provide.
• Brainstorm: suggest possibilities for a person to evaluate.

A vague request such as "Help me with marketing" leaves many decisions open. A clearer request might be:

Create a one-week content outline for a fictional home-services company. The audience is homeowners who need routine maintenance. Suggest one educational post and one customer-question post for each weekday. Do not write advertisements that promise a particular result.

Now the model knows the task, audience, time span, and an important boundary.

Describe the finish line

A useful prompt explains what success looks like. Consider these questions:

• What should the response help you accomplish?
• Who will read or use it?
• What information must be included?
• What must be left out?
• How will you decide whether the result is acceptable?

For a customer email, success might mean that the answer addresses the customer's question, uses the approved policy, sounds respectful, and leaves out any fee that has not been confirmed.

For a meeting summary, success might mean that decisions, open questions, owners, and next steps are separated rather than blended into one paragraph.

Separate the main task from optional work

If you ask for a customer email, a marketing plan, a legal review, and a spreadsheet in one request, the model may produce a broad response that does none of those jobs particularly well.

Break complicated work into manageable stages. First ask for an outline. Review it. Then ask for a draft of the first section. This keeps the task easier to inspect and gives you a chance to correct assumptions before the work grows.

3. Give the model relevant context

Context is the background information that helps the model understand the request. It can include the audience, situation, purpose, definitions, earlier decisions, and relevant constraints.

Without context, the model may fill gaps with generic assumptions. With relevant context, it has more information to use.

For example, "Write an appointment reminder" does not tell the model whether the message is for a first-time customer, whether the appointment is confirmed, or what tone the company prefers.

A more complete request might say:

The customer has already confirmed a maintenance appointment. Write a friendly reminder that asks them to contact the office if they need to reschedule. Do not add an appointment time because it is not included in the information below.

The prompt gives the model the situation and explains what is missing.

Add the information that changes the answer

Useful context often includes:

• Audience: who will read the result.
• Purpose: what the result should accomplish.
• Situation: what has happened so far.
• Definitions: what a term means in this particular task.
• Preferences: tone, reading level, or house style.
• Relevant history: earlier decisions that still apply, if they are available and appropriate to share.

Do not add background merely to make the prompt longer. Include details that change what a good answer should look like.

A role can help, but it does not create expertise

You may begin with a role such as:

You are a writing assistant helping a small service company prepare clear customer communications.

That gives the model a useful perspective for the task. It does not make the model a licensed technician, attorney, accountant, or employee with access to company records. A role description changes the framing of the request; it does not grant credentials, authority, or access.

Use only information you are permitted to share

Context can contain sensitive information. Follow your organization's rules and the AI provider's privacy terms. Remove names, addresses, account details, payment information, health information, and other personal data unless the tool and workflow have been specifically approved to handle it.

A better prompt cannot make an unapproved data-sharing practice safe.

4. Provide source material and tell the model how to use it

A prompt can tell the model to use a particular source, but the source itself must be available in the context or through an authorized tool.

This is especially important when an answer depends on a current business fact. The model's general language ability does not mean it knows a company's latest prices, policies, appointment availability, inventory, or customer records.

Imagine Harborlight receives this question:

Do you charge an extra fee for weekend appointments?

For this exercise, Harborlight's fictional approved FAQ says:

The office can explain available appointment options. The FAQ does not list a weekend service fee. Staff must confirm any fee with the office before quoting it.

A weak prompt would be:

Answer the customer's question about weekend fees.

A stronger prompt would identify the FAQ as the source and explain what to do if the answer is absent.

Label the source

When you paste reference material, give it a clear heading such as Approved FAQ, Current policy excerpt, or Meeting notes. This helps distinguish your instructions from the material being discussed.

Then state the source rule:

Use only the approved FAQ below for policy details. Do not add a fee or infer one. If the FAQ does not answer the question, say that the information is not provided and identify what a staff member must confirm.

Ask for a boundary between known and unknown

Useful instructions include:

• Use only the supplied source for policy details.
• Do not fill gaps with assumptions.
• If a requested fact is missing, say that it is missing.
• If two supplied sources conflict, identify the conflict instead of choosing one silently.
• Do not invent a citation, policy clause, date, name, or amount.
• Separate facts from suggestions.

These instructions can reduce unsupported additions, but they do not prove that the model followed them. Compare the answer with the source before relying on it.

Source material is not the same as live access

Pasting a policy gives the model that pasted text. It does not automatically connect the model to the company's policy database or guarantee that the pasted version is current. If a workflow uses an authorized connection, confirm what information that connection actually provides and whether it is current.

The basic rule is simple: if a fact matters, provide an approved source or retrieve it through an approved system, then verify the result.

5. Set boundaries and explain what to do when information is missing

A good prompt describes not only what the model should do, but also what it should avoid.

These instructions are often called constraints or guardrails. They define the boundaries of the requested response.

Examples:

• Do not invent prices, dates, policies, or names.
• Do not promise an outcome that the source does not support.
• Do not include personal information.
• Do not change the meaning of the supplied policy.
• If a required detail is missing, ask a question instead of guessing.
• If the source is unclear, identify the unclear passage.

Make uncertainty actionable

"Be accurate" is a reasonable goal, but it does not tell the model what to do when the answer is uncertain.

Try a specific instruction:

If the supplied material does not answer the question, write: "The supplied information does not confirm this." Then list the exact detail that needs human confirmation. Do not create a likely answer.

This gives the model a response pattern for missing information.

Distinguish hard controls from written instructions

A prompt can ask a model not to perform an action. It is not the same as a permission system that technically prevents that action.

For example, telling a model not to send an email is not a substitute for an application that requires human approval before sending. Telling a model not to change a price is not a substitute for software permissions that prevent unauthorized price changes.

Use prompts to communicate the task and boundaries. Use application controls, validated data, and human approval for actions that must be restricted.

Avoid contradictory instructions

A prompt can accidentally ask for incompatible things:

Explain every detail, but keep the answer to one sentence.

Or:

Use only the supplied FAQ, and add current market prices from the internet.

When instructions conflict, the model may choose one, blend them, or ask for clarification. Write priorities explicitly:

Use the supplied FAQ for policy details. If a question requires information outside the FAQ, identify what is missing. Do not search for or invent a replacement policy.

Clear priorities make the request easier to follow.

6. Specify the audience, tone, length, and format

A useful answer is not just correct in substance. It also needs to be usable by the person who receives it.

Tell the model who the response is for and how it should be presented.

Audience and reading level

Compare:

Explain the service process.

With:

Explain the service process to a homeowner who has never scheduled a maintenance visit. Use everyday language and explain any technical term the first time it appears.

The second prompt tells the model what the reader already knows and what may need explanation.

Tone

Tone words are helpful, but examples are often clearer. You might ask for a message that is:

• Friendly and professional.
• Calm and reassuring without making promises.
• Direct and respectful.
• Warm but not overly casual.
• Clear without sounding defensive.

If a particular phrase or style is important, show an approved example and explain what should be retained.

Length

Be specific about the intended size:

• A short text message.
• A three-paragraph email.
• A one-page outline.
• A checklist with no more than a few words per item.

A word limit is a target, not a guarantee. Review the result and shorten it if necessary.

Format

Name the structure you want:

Separate the draft from the review notes

For a business task, you might ask:

First provide the customer-facing draft. After it, add a separate section titled "Staff review" listing any fact, amount, date, or policy detail that the source did not confirm. Do not include the staff review section in the customer message.

This makes it easier for a person to see what is ready to edit and what still needs attention.

7. Use examples to show the pattern you want

Sometimes a description is not enough. You can include an example of the kind of input and output you want the model to imitate.

This technique is often called few-shot prompting. The examples show a pattern in context. The model can use that pattern when producing a new response.

For example, you might provide:

Example input

Customer: I need to move my appointment. What should I do?

Example output

Thanks for letting us know. Please contact the office so the team can review the available appointment options with you.

Then ask the model to use the same style for a different customer question.

Make examples representative

A useful example should demonstrate the features that matter:

• The level of detail.
• The tone and vocabulary.
• The structure.
• The difference between confirmed information and information that needs review.

If every example is unusually short, the model may produce answers that are too brief. If an example includes a fact that should not be repeated, the model may copy it into a new answer.

Include a counterexample when it helps

You can show what not to do:

Do not promise that a technician will arrive at a particular time unless the approved schedule confirms it.

A counterexample can clarify a boundary, especially when the mistake is easy to make.

Examples are guidance, not guarantees

The model may imitate an example too closely, copy an irrelevant detail, or miss the pattern you intended. Use fictional or approved sample data. Remove personal information. Review the result to make sure the model followed the intended pattern rather than copying the example's facts.

Start with a small number of clear examples. Add more only when they teach a distinct pattern that the model needs to recognize.

8. Build a prompt using a repeatable framework

You do not need a complicated formula for every request. A simple framework can help you remember the pieces that matter.

The six-part prompt plan

Purpose: What job should the model perform?

Context: What background does it need to understand the situation?

Source: Which supplied or authorized information should it use?

Boundaries: What must it avoid, and what should it do when information is missing?

Output: Who is the audience, and what format, tone, and length do you want?

Review: What should the model flag for a person to check?

Not every prompt needs all six parts. A simple request to explain a common term may need only a task and audience. A customer response involving a company policy needs more.

Reusable prompt template

Copy this structure and replace the bracketed instructions with information appropriate to your task:

You are helping with [type of task].

TASK
[Describe exactly what you want produced.]

CONTEXT
[Explain the situation, audience, and purpose.]

SOURCE
[Paste or identify the approved source material.
State which source controls if more than one is supplied.]

BOUNDARIES
[State what not to invent, change, disclose, or assume.
Explain what to do when information is missing.]

OUTPUT
[Specify tone, format, length, and required sections.]

REVIEW
[List facts or decisions a person must verify before using the result.]

A completed fictional example

You are helping a fictional home-services company draft customer replies.

TASK
Draft a reply to a customer asking whether a weekend service fee applies.

CONTEXT
The customer needs a clear answer. The reply should be polite and concise.

SOURCE
Use only the fictional approved FAQ supplied below:
"The office can explain available appointment options.
The FAQ does not list a weekend service fee.
Staff must confirm any fee with the office before quoting it."

BOUNDARIES
Do not invent a fee or imply that the fee is zero.
If the FAQ does not answer the question, say that the office
must confirm the information.

OUTPUT
Write a short customer-facing email in a friendly,
professional tone.

REVIEW
After the email, list any fact that still needs staff confirmation.

The completed example gives the model a task, context, source, boundaries, output expectations, and a review requirement. It does not make the model the authority on the company's fee policy. A person still compares the draft with the approved source before sending it.

9. Improve prompts through deliberate iteration

Your first prompt is a starting point. Prompt engineering usually involves trying a request, inspecting the response, and making a purposeful change.

Avoid changing everything at once. If you change the audience, source, tone, and format together, it becomes difficult to tell which change improved the result.

A practical improvement loop

1. Write the first prompt. State the task and the information available.
2. Run it. Read the response instead of judging it by how polished it sounds.
3. Identify the problem. Is it inaccurate, incomplete, too long, too vague, or in the wrong format?
4. Change one important part. Add missing context, clarify the source, or change the output requirement.
5. Run it again. Compare the new response with the original.
6. Check against the success criteria. Keep the change only if it improves the result without creating a new problem.

Ask for a diagnosis, not just another attempt

Instead of saying:

Try again.

You can say:

The draft added a weekend fee that does not appear in the supplied FAQ. Revise the prompt so the response must identify missing fee information rather than guessing. Then produce a new draft using only the FAQ.

This identifies the failure and the behavior you want changed.

Test more than one example

A prompt that works for one easy question may fail when the source is incomplete or the customer asks something unexpected.

Try a small set of different cases:

• The source clearly answers the question.
• The source does not contain the requested fact.
• Two supplied passages appear to conflict.
• The customer asks for something outside the task.
• The source contains a phrase that could be misunderstood.

Look for repeated problems. A reusable prompt should work acceptably across the kinds of requests it is meant to handle, not only the example used to create it.

Save useful prompts with their purpose

For repeated work, keep an approved copy of the prompt, the intended source, and a few test examples. Give it a clear name and a version date if your process requires one. When you change it, test it again.

This turns prompt improvement into a repeatable work process rather than a collection of random phrases.

10. Know what prompt engineering cannot do

Prompt engineering is useful, but it has limits.

A prompt cannot automatically:

• Give a model access to a file, database, website, or live schedule that the application has not provided.
• Make an outdated source current.
• Turn an uncertain answer into a verified fact.
• Guarantee that the model will follow every instruction.
• Grant professional credentials or decision-making authority.
• Replace software permissions, validation, or required human approval.
• Make it appropriate to share confidential information with an unapproved service.

A role such as "act as an expert" can influence the style and framing of a response. It does not prove that the model has the qualifications or information of that expert.

A request such as "never make a mistake" does not create a technical guarantee. For important tasks, use an approved source, a suitable system of record, ordinary software checks where possible, and a qualified person to review the result.

Use the right tool for the job

If you need an exact calculation, use a calculator or tested software and verify the inputs. If you need a current appointment slot, use the authorized scheduling system. If you need a company policy, consult the approved policy source. A language model may help explain, summarize, or draft around that information, but it should not replace the source of truth.

Keep private information private

Before entering information into an AI tool, check whether the tool is approved for that information and what its privacy and data-handling terms say. Use fictional or de-identified examples when practicing. Do not paste customer records, payment details, passwords, private health information, or confidential business information into an unapproved tool.

Keep a person responsible

A prompt can help prepare a draft. A person remains responsible for checking the facts, deciding whether the response is appropriate, and approving consequential actions. This is especially important when a response affects safety, money, legal rights, employment, or a customer's account.

The goal is not to make the model sound certain. The goal is to make its work easier to inspect and safer to use.

11. What to remember

A strong prompt gives the model a clear job and the context needed to attempt it.

Remember these principles:

• State the task in plain language.
• Explain the audience and purpose.
• Supply relevant context instead of expecting the model to know your situation.
• Identify the approved source for facts that matter.
• Tell the model what to do when information is missing.
• Specify tone, format, and length when they affect the result.
• Use examples when you need to demonstrate a pattern.
• Improve prompts by testing and changing them deliberately.
• Verify important facts and keep a person responsible for decisions.

Lesson 4 explained why prompts matter: the model predicts likely next tokens from the context it receives. Lesson 5 has shown how to shape that context with care.

The best prompt is not necessarily the longest prompt. It is the prompt that gives the model enough relevant information to do a clearly defined job, while making the result easy for a person to review.