3

Class 3: Prompting Well

Class 3 of 8About 3 hoursNo coding

By the end of this class you will be able to

• Write a prompt using the six building blocks: role, task, context, format, examples and constraints.
• Improve a weak answer through follow-up messages instead of starting again from scratch.
• Use examples to get the tone and layout you want (few-shot prompting).
• Ask the AI to work step by step, and to ask you questions before it drafts.
• Avoid the most common prompting mistakes, including pasting data you are not allowed to share, and keep a reusable prompt template for your own work.

Where we are

In Class 2 you saw that a chatbot predicts the next word based on the text it is given. This class is about writing that text well. Prompts are the first part of the harness, the instructions and structure wrapped around the model, on the AI Map. In Class 4 we look at giving the AI the right information to work with.

1. Why prompts matter

Imagine a very capable new colleague on their first morning. They have read a huge amount, they write quickly, they never get tired and they are keen to help. But they have never been to your office. They don't know your programme, your donors, your acronyms, who reads your reports, or what "the usual format" means. They also won't ask many questions. Give them a vague request and they will do their best, with confidence, and fill the gaps with sensible-sounding guesses.

That is roughly what you are working with every time you open an AI chat tool.

If you say to this new colleague "Write something about the floods", you will get something. It might be a news article, a poem or a history of flooding. It will probably be generic, too long and pitched at the wrong reader. That isn't their fault. You didn't tell them what you needed.

Now imagine you say: "You're helping me prepare the weekly situation report for our country director. Here are this morning's field notes from Aramu district. Pull out the number of people affected, the main needs and what's already being done, in five bullet points of one line each. If a figure isn't in the notes, write 'not stated'. Don't guess." Same colleague, very different result.

A prompt is simply everything you type into the AI tool: your question, your instructions, any text you paste in, and your follow-up messages. Prompting is the skill of writing those instructions so the AI produces something useful. It isn't programming and it isn't a secret code. It is the same skill as briefing a person well, and most professionals already have it. You write terms of reference, brief consultants and give handover notes. This class shows you how to use that skill with AI.

The one-sentence version

The AI can only use what you give it. Everything it doesn't know about your situation it will guess, and it guesses in fluent, confident prose.

Remember from Class 2 that the model works by predicting likely next words given the text in front of it. A short, vague prompt gives it very little to go on, so it falls back on the most average answer, like a generic essay anyone might have asked for. A detailed prompt pulls the prediction towards your situation.

2. The anatomy of a good prompt

Good prompts tend to contain the same six building blocks. You won't need all six every time. A quick question needs one or two. But when an answer disappoints you, the missing block is nearly always on this list.

Block	What it tells the AI	Example phrase
Role	Who it should act as, which sets expertise, vocabulary and viewpoint	"You are an experienced humanitarian reporting officer."
Task	What exactly you want done, using a clear action verb	"Summarize the field notes below into a situation update."
Context	The background: who it's for, why, what's already known	"The reader is the country director, who has two minutes and needs to decide on funding today."
Format	The shape of the answer: length, structure, headings, tables	"Use five bullet points, one line each, then one line headed 'Gaps'."
Examples	A sample of what good looks like	"Here is last week's update. Match its style."
Constraints	Rules and limits: what to avoid, what to do when unsure	"Use only information in the notes. If a figure is missing, write 'not stated'."

A few notes on each:

• Role is the least important of the six, though people often start with it. "You are an expert…" helps a bit, mostly by setting vocabulary and tone. It does not give the AI knowledge it doesn't have. A role without a clear task and context achieves little.
• Task should use a precise verb. "Summarize", "compare", "rewrite for", "list", "extract", "draft", "critique" and "translate" each lead to different outputs. "Help me with" or "do something about" leaves the AI to guess.
• Context is where most of the value is, and it's what people leave out most. The AI can't see your inbox, your programme documents or your manager's preferences unless you tell it. Who is the audience? What will they do with this? What happened last time?
• Format saves you editing time. If you need 150 words, say 150 words. If you need a table with specific columns, name the columns.
• Examples are so powerful they get their own section below.
• Constraints protect you. The most useful one for professional work is "If the information isn't in what I gave you, say so. Don't make it up." Class 2 explained why chatbots invent things. This constraint doesn't eliminate that risk, but it reduces it a lot.

Know the Term: System prompt

A system prompt is a set of standing instructions given to the AI before the conversation starts. It usually sets the AI's role, its rules and its style, such as "You are an assistant for humanitarian staff. Answer in plain English. Never invent figures." The people who set up the tool write it, and you usually cannot see it. Many chat tools also let you write your own version, often called custom instructions, a project, or a custom assistant, that applies to every chat in that space. Your prompt is added on top of it.

Analogy: the standing orders a new colleague receives in their induction pack. Each request you make is a new task, but the standing orders still apply.

Why you care: if you retype the same background ("I work in programme support, write in British English…") in every chat, put it in custom instructions once. It also explains why the same question can get different answers in different tools.

Weak prompt vs strong prompt

Here is the same request written two ways. Both are about a (fictional) flood in Northern Veloria province.

Weak prompt

Write a sitrep about the floods in Veloria.

What you get: a confident, generic report that may include invented figures, invented organizations and a made-up timeline. The AI knows nothing about your floods, so it writes about floods in general. It doesn't know who the reader is, how long it should be or what format your office uses.

Strong prompt

You are a humanitarian reporting officer. [ROLE]

Turn the field notes below into a short situation update. [TASK]

The update is for our country director, who will read it on her phone before a 9:00 funding meeting. She needs to know how many people are affected, the top needs, and what is already being done. [CONTEXT]

Format: a one-line headline, then these headings: Affected population / Priority needs / Response so far / Gaps and unknowns. Under 200 words in total. [FORMAT]

Use only the information in the notes. If a figure is missing or unclear, write "not stated" or "unclear" — do not estimate. Flag any dates that could be read two ways. [CONSTRAINTS]

Field notes:
"""
(paste the notes here)
"""

What you get: a short, structured update built from your information, with the gaps clearly marked for you to follow up.

You don't need the labels in square brackets. They are only there to show the building blocks. Notice also the triple quotation marks (""") around the pasted notes. Marking clearly where your instructions end and your pasted material begins helps the AI keep them apart. Any clear marker works: quotation marks, a line of dashes, or a heading such as "Field notes:".

Longer isn't automatically better

The strong prompt is longer, but its length isn't what makes it strong. Every sentence gives the AI something it couldn't have guessed. Padding such as long preambles, repeated instructions or three paragraphs of background it doesn't need can bury the important part. Aim for complete, not long.

Know the Term: Structured output

Structured output means asking the AI to answer in a fixed, predictable shape instead of free-flowing text: a table with named columns, a form with set headings, or a data format such as JSON (a plain-text format that computer programs can read, made of labelled fields like "district": "Aramu"). You describe the shape, meaning the fields, their order and what to write when something is missing, and the AI fills it in. Software that uses AI goes a step further: it can require the model to follow an exact template and reject any answer that does not fit.

Analogy: handing someone a printed form to fill in instead of a blank sheet of paper. Every answer lands in the same box, every time.

Why you care: structured answers are faster to check, easy to compare across reports and ready to paste into a spreadsheet or another system. An empty box also shows you straight away what information is missing.

3. Iterating: a conversation, not a vending machine

A vending machine works like this: you press a button, something drops out, and if it's wrong you walk away. Many people use AI tools this way. They type one prompt, look at the answer, decide "AI isn't good at this", and close the window.

A chat tool is a conversation. The first answer is a first draft. The real skill is often in the second, third and fourth messages. Within a single chat, the AI can see everything said so far, so you can build on it instead of starting again.

The AI only remembers this conversation

The AI can see the messages in the current chat. In most enterprise tools, a new chat starts from a blank page. It doesn't remember what you told it yesterday unless your tool has a clearly labelled memory or "project" feature switched on. If you start a new chat, give the context again.

Here are four follow-up moves that work.

1. Tell it exactly what to fix. "That's not quite right" gives the AI very little to work with. Be specific, as you would with a colleague's draft:

Good start. Three changes:
1. Cut it to half the length.
2. The reader is the donor, not our own team — remove internal acronyms.
3. Move the water situation to the top; it is the most urgent need.

2. Ask it to ask you questions first. When you're not sure what context the AI needs, let it tell you. This is one of the most useful habits in this class:

I need to write a two-page concept note for a hygiene project in Aramu district. Before you draft anything, ask me the questions you need answered to do this well. Ask them one at a time.

The questions themselves are often useful, because they show you what a good concept note needs and what you hadn't yet thought about.

3. Ask for options, not one answer. The AI can produce alternatives in seconds. Take advantage of that:

Give me three different subject lines for this email: one formal, one warm, one very short. Tell me which you would pick and why.

4. Ask it to critique its own work. A second look often catches problems:

Now review your summary against the original notes. List any figure, name or date in your summary that does not appear in the notes.

This last move is worth making a habit for anything that includes numbers. It isn't a replacement for checking yourself (Class 6 covers verification in depth), but it catches a surprising number of slips.

When to start a fresh chat

If a conversation has gone badly off course, or has become very long and the AI seems to be "forgetting" your early instructions, start a new chat. Paste in your improved prompt, written with everything you learned from the failed attempt. You're not starting over, you're starting with a better brief.

4. Showing, not telling: examples (few-shot)

Some things are hard to describe but easy to recognise: a tone of voice, a house style, the right level of detail. "Write it in a warm but professional tone" means different things to different people. Showing an example removes the guesswork.

Giving the AI one or more samples of what you want is called few-shot prompting. With no examples it's called zero-shot. It works because the model is very good at continuing patterns. Show it two examples that follow a pattern and it will follow the same pattern for the third.

Know the Term: Zero-shot vs few-shot prompting

A zero-shot prompt asks the AI to do a task with instructions only and no examples: "Write a one-line summary of this report." A few-shot prompt adds a handful of worked examples (the "shots") of the input and the output you want, then gives the new input. The model spots the pattern in your examples, such as the length, tone, structure and level of detail, and copies it. Two or three good examples often do more than a paragraph of description, because style is easier to show than to explain.

Analogy: handing a new colleague two past donor letters and saying "one like these", instead of writing a page of style guidance.

Why you care: when the AI keeps getting the tone or layout "almost right", examples are usually the quickest fix. Just make sure they are good ones, because the AI will copy their flaws too.

Here is a few-shot prompt for turning dry figures into short updates for a donor newsletter:

Turn each set of facts into a two-sentence donor update. Follow the style of the examples exactly.

Facts: 300 hygiene kits distributed in Lomi village; funded by the Northern Fund. Update: Thanks to the Northern Fund, 300 families in Lomi village now have soap, water containers and purification tablets. These kits are a first line of defence against waterborne disease after the floods.

Facts: Mobile clinic reopened in Tavet; 85 patients seen in first week. Update: The mobile clinic in Tavet is open again, and 85 people were treated in its first week. For many families it is the only health service within a day's walk.

Facts: Temporary classrooms set up for 240 children in Aramu town. Update:

The AI will almost certainly produce two sentences, lead with the impact on people, and end with why it matters, because that's what both examples did. You never had to explain "lead with people, then give significance". The examples said it for you.

Tips for good examples:

• Two or three is usually enough. One example can be copied too literally. Several varied examples show the pattern rather than one instance of it.
• Vary the content, keep the style. If both examples are about hygiene kits, the AI may assume every update is about hygiene kits.
• Your examples set the standard. If your sample contains jargon, a typo or an unsupported claim, expect the AI to reproduce it.
• Real past work makes good examples, as long as you are allowed to share it with the tool. An approved, published report from last year is fine. An internal draft with personal details is not.

5. Asking the AI to think step by step

Chatbots write one word at a time, and each word is influenced by the words already written. So if the AI jumps straight to an answer, it has had no chance to "work anything out" first. When a task involves several steps, such as adding up figures, comparing options or checking a list of conditions, asking the AI to reason through it step by step before answering tends to give better results. This is called chain-of-thought prompting.

Know the Term: Chain of thought

Chain of thought means getting the AI to work through a problem in visible steps before it gives its final answer, for example by adding "Think step by step" or by listing the steps you want it to follow. Because a chatbot writes one word at a time, the words it has already written guide what comes next. Writing out the intermediate steps gives it more to build on, which tends to improve answers on tasks involving counting, comparing, sums or several conditions. Many newer reasoning models do this thinking on their own before they answer.

Analogy: a maths teacher who says "show your working". The working helps the student get it right, and it lets the teacher see where it went wrong.

Why you care: asking for the steps makes mistakes easier to spot. But the steps can still be wrong, and a tidy explanation is not proof. Check the numbers that matter.

The simplest version is to add "Think this through step by step before giving your final answer." Even better is to spell out the steps yourself, because then you control the method:

Three field teams sent the figures below for displaced households in Aramu district. Before giving a total:
1. List each team's figure and which villages it covers.
2. Point out any village counted by more than one team.
3. Point out any figure that is a range or estimate rather than a count.
4. Only then give a total, and say how confident you are in it and why.

Team A: Aramu town, 640 households.
Team B: Lomi and Tavet, about 350–400 households.
Team C: Tavet and Dorra, 210 households (Dorra not visited, estimate from village leader by phone).

A one-shot "what's the total?" might well return a neat single number. The step-by-step version should show you that Tavet may be counted twice, that Team B gave a range, and that Dorra's figure is second-hand. Those are exactly the things a careful reporting officer would flag.

Two cautions:

• Visible reasoning isn't proof. A well-organised explanation can still contain a wrong number. The value is that the steps make errors easier to find, so read them.
• Some tools already do this. Newer "reasoning" models think before answering on their own (as of 2026). Laying out the steps you want is still useful, because it makes the AI follow your method and show its working in a way you can check.

6. Common mistakes

Mistake	Why it fails	Fix
Being vague ("Write about the floods")	The AI fills every gap with generic guesses	Name the task, audience, purpose and format
Leaving out context you think is obvious	The AI can't see your office, your files or your history	Imagine briefing someone on their first day. What would they need to know?
Asking for everything at once (report + email + slides + translation)	Quality drops, and the AI skips or blends parts	Break the job into steps and do one at a time in the same chat
Believing the first answer	First drafts contain errors and invented details	Iterate, ask it to check itself, and verify facts and figures yourself
Assuming it remembers earlier chats	Most tools start each new chat from zero	Re-paste key context, or use your tool's approved custom instructions or project feature
Shouting or pleading ("YOU MUST!!!", "please please")	Capital letters and politeness don't unlock better answers. Clarity does	Say what you want plainly and explain why it matters
Padding the prompt	Important instructions get lost in long, repetitive text	Every sentence should tell the AI something it couldn't guess
Not saying what to do when information is missing	The AI tends to fill gaps with plausible guesses	Add: "If it isn't in the text, write 'not stated'. Don't estimate."
Pasting confidential or personal data	Information you paste may be stored, logged or seen outside your team, depending on the tool and its approval	Use only approved tools, only for data they are approved for, and remove names and identifying details

Before you paste anything

Before you paste a document into an AI tool, ask yourself: Is this tool approved for this kind of information? Beneficiary names, medical details, staff matters, security information and unpublished negotiating positions need particular care. When in doubt, anonymise ("Village A", "Staff member 1") or don't paste it. See the data rules on the About page. Class 8 covers data safety in full.

A word on politeness. Saying "please" and "thank you" is harmless, and it is fine if it keeps your writing natural. But it isn't a technique. The answer improves because of the content of your request, not its manners. The same goes for capital letters, exclamation marks and threats. If the AI keeps ignoring an instruction, the usual reason is that the instruction is unclear, conflicts with another one, or is buried in the middle of a long prompt. Move it to the end, state it plainly and give the reason: "Keep it under 100 words, because it will be read aloud in a two-minute radio slot."

7. Reusable prompt templates

Once a prompt works well, don't lose it. Many of your tasks recur: the weekly update, the meeting summary, the email declining an invitation, the translation check. A saved template turns a good prompt into a habit and saves you rethinking the brief every time.

Here is a general template built from the six blocks. Copy it into a document you keep handy, fill in the square brackets and delete any line you don't need.

ROLE: You are [a role with relevant expertise, e.g. "an experienced programme officer in a humanitarian organization"].

TASK: [Action verb] [what you want], e.g. "Draft a reply to the email below, accepting the meeting and proposing two alternative times."

CONTEXT:
- This is for [who will read it] who needs to [what they will do with it].
- Background: [what the AI needs to know that it can't guess].
- The tone should be [e.g. formal / warm / neutral / urgent].

FORMAT: [length, structure, headings, bullets or table columns], e.g. "Under 150 words, three short paragraphs."

EXAMPLES (optional): Here is an example of the style I want:
"""
[paste a good past example]
"""

CONSTRAINTS:
- Use only the information I provide. If something is missing, say
  "not stated" instead of guessing.
- Avoid [jargon / acronyms / specific words or claims].
- [Any other rule, e.g. "British English", "no bullet points"].

Before you start, ask me up to three questions if anything important is unclear.

MATERIAL:
"""
[paste the text to work on, after checking it is safe to share]
"""

Some ways to get more out of templates:

• Build a small library. Three to five templates for your most frequent tasks is plenty. Name them clearly, for example "Weekly SitRep", "Donor update", "Meeting summary".
• Share them with your team. A shared template gives more consistent results across colleagues. It is also a quiet way to build in good habits such as "write 'not stated'".
• Improve them over time. When you have to correct the same thing twice, add a line to the template so you never have to correct it again.
• Use your tool's features if they are approved. Many tools let you save standing instructions (see the system prompt sidebar above) so your background context applies automatically.

Prompt Lab

Work in pairs. Use your organization's approved AI chat tool. All the exercises use the fictional field report below, from our running case study, the SitRep Assistant. Copy the report once and keep it ready to paste.

The field report (fictional)

FIELD NOTE - Aramu district, Northern Veloria - sent Thurs eve From: K. Osei, Veloria Relief Coalition field team

heavy rain since the weekend, Tula river burst near Kessa bridge. we got to Aramu town + Lomi and Tavet villages today. could NOT reach Dorra, road cut after the bridge, no phone signal there either.

approx 1,200 households displaced in total (our estimate + figures from village chiefs). most are staying at Aramu primary school, which is full - maybe 600 people there. more families in the community hall but we didn't manage to count them.

water is the big problem. 2 of the 3 boreholes in Lomi are flooded and people are drinking from the river. Tavet health post still open but running low on ORS, nurse reports 14 cases of acute diarrhoea since Monday, mostly children under 5. no deaths reported to us.

VRC is distributing food at the school, enough stock for about 3 days. next truck expected 04/05 if the Kessa road is passable. urgent: water purification, hygiene kits, ORS, shelter materials for the hall.

Notice what's messy about it: the date is only "Thurs", the next delivery date "04/05" could be 4 May or 5 April, the number of people in the community hall is missing, and "1,200 households" is partly an estimate. A good SitRep flags all of this instead of hiding it.

Exercise 1: From weak to strong

Task: First, send the weak prompt with the report pasted underneath. Then open a new chat and send the strong version. Compare the two answers side by side.

Summarize this.

You are a humanitarian reporting officer. Turn the field note below into a short situation update for our country director, who will read it on her phone before a funding meeting and needs to know: how many people are affected, the most urgent needs, and what is already being done.

Format: a one-line headline, then four headings: Affected population / Priority needs / Response so far / Gaps and unknowns. Under 200 words.

Use only information in the note. Do not estimate missing figures. If a date could be read two ways, say so.

Field note:
"""
[paste the field report here]
"""

What to look for:

• Did the weak version invent or "tidy up" anything, such as a total for the community hall or a definite date?
• Does the strong version put the most urgent need (water) near the top?
• Is "04/05" flagged as ambiguous, and is Dorra listed as a gap?
• Which of the six building blocks made the biggest difference, in your view?

Exercise 2: A table, with honest gaps

Task: Ask for structured output, with a fixed set of columns and a clear rule for missing information.

From the field note below, extract the key facts into a table with these columns: Location | People or households affected | Main needs | Response under way | Source or confidence.

Rules:
- One row per location mentioned (Aramu town, Lomi, Tavet, Dorra, plus
  one row for the community hall if it is described separately).
- If information is not in the note, write "not stated". Never estimate.
- In the last column, say whether each figure is a count, an estimate,
  or second-hand.

After the table, list every date in the note and explain any that are ambiguous.

Field note:
"""
[paste the field report here]
"""

What to look for:

• Does Dorra's row say "not stated" instead of made-up figures?
• Is the community hall number left blank ("not stated") rather than guessed?
• Does it separate the 1,200 households from the ~600 people, or wrongly mix the two units?
• Would this table be easy to paste into a spreadsheet and compare with tomorrow's report?

Exercise 3: Matching a tone with examples

Task: Use few-shot prompting to turn the report into a short donor update that matches an existing house style.

Write a short update for our donors' newsletter based on the field note below. Match the style of these two past updates exactly: length, tone and structure.

Example 1:
"When the Sela river flooded in March, 400 families lost their homes overnight. Within 48 hours, your support put clean water and hygiene kits in their hands. The work is not over: 90 families are still waiting for shelter materials."

Example 2:
"In Kopa district, a single health post serves 12,000 people. Thanks to you, it now has a month's supply of essential medicines. Next, we aim to train two more community health workers."

Rules: no figures that are not in the note; do not name individual staff; do not mention exact locations of shelters.

Field note:
"""
[paste the field report here]
"""

What to look for:

• Does the update follow the pattern of the examples: people first, then the support, then what's still needed?
• Is it about the same length as the examples?
• Did it keep to the rules: no staff names ("K. Osei"), no invented figures?
• Try it again with no examples (zero-shot). What changes?

Exercise 4: Let the AI interview you

Task: Instead of writing a full brief, let the AI find out what it needs to know. One partner plays the "programme officer" and answers its questions. Make up plausible fictional answers.

I need to draft a one-page funding appeal for the Aramu flood response, based on the field note below. Before you write anything, interview me:
ask the questions you need answered to write a strong appeal, one at a time, and wait for my answer before asking the next. When you have enough, tell me, summarize what you learned, and then write the draft.

Field note:
"""
[paste the field report here]
"""

What to look for:

• Were the questions useful? Did any of them point out something you hadn't considered, such as the target donor, the amount needed or the deadline?
• Did the final draft use your answers, or drift back to generic text?
• Did it ask about the gaps in the note (Dorra, the community hall)?
• Would you use this "interview me" move in your own work? For which tasks?

Exercise 5: Bring your own task

Task: Pick a real, recurring task from your own job that is safe to use with your approved tool (no confidential or personal data). Fill in the template and try it. Then improve it with at least two follow-up messages.

ROLE: You are [role].

TASK: [Action verb] [what you want].

CONTEXT:
- This is for [reader] who needs to [purpose].
- Background: [what the AI can't guess].
- Tone: [tone].

FORMAT: [length and structure].

EXAMPLES (optional):
"""
[a good past example]
"""

CONSTRAINTS:
- Use only the information I provide; write "not stated" if something
  is missing.
- [other rules]

Before you start, ask me up to three questions if anything important is unclear.

MATERIAL:
"""
[your material, checked as safe to share]
"""

What to look for:

• Which block did you find hardest to fill in? That's often the one your past prompts were missing.
• What did you have to correct in your follow-ups? Add a line to the template so you won't need to next time.
• Save the finished template somewhere you'll find it next week.

Check Your Understanding

1 You ask an AI tool to "Write a sitrep about the floods" and get a confident, generic report with figures you don't recognise. What is the most likely cause?

1. The AI tool is broken or low quality and should not be used for reporting.
2. The prompt gave no context, so the AI filled the gaps with generic guesses.
3. The prompt was not polite enough; adding "please" would have produced real figures.
4. The AI needs to search the internet first; once it can, prompts don't matter.

2 Which of the six building blocks do people most often leave out, even though it usually adds the most value?

1. Role ("You are an expert…")
2. Format (length, headings, bullets)
3. Context (who it's for, why, and the background the AI can't know)
4. Capital letters for the most important instructions

3 Yesterday you spent 20 minutes explaining your programme to the AI in a chat. Today you open a new chat and ask for "the usual weekly summary". What should you expect?

1. It probably won't know your programme. Unless your tool has a memory or project feature switched on, a new chat starts from zero.
2. It will remember everything, because the AI learns from each conversation.
3. It will remember, but only if you ask politely.
4. It will refuse to answer until you open yesterday's chat.

4 The AI keeps writing donor updates that are accurate but in the wrong tone: too formal and too long. What is usually the quickest fix?

1. Make the prompt much longer, with several paragraphs describing the tone.
2. Write the instruction in capital letters: "BE WARM AND SHORT!!!"
3. Start a new chat and send the same prompt again until the tone comes out right.
4. Paste two or three past updates you like and ask it to match their style (few-shot).

5 You ask the AI to total displaced households from three teams' figures. Some villages appear in two teams' reports. Which prompt is most likely to give a result you can check?

1. "What's the total? Just the number, no explanation."
2. "List each team's figure and villages, flag any village counted twice and any estimates, then give the total."
3. "You are the world's best statistician. Give me the total."
4. "Give me the total. Be 100% accurate."

6 The AI's first draft of a SitRep is about 70% right. What is the best next step?

1. Give up; if the first answer isn't right, AI isn't suited to this task.
2. Reply "That's wrong, try again."
3. Reply with specific changes, e.g. "Move water needs to the top, cut to 150 words, and mark the Dorra figures as 'not stated'."
4. Copy the draft into the final report as-is; you can correct it later if anyone complains.

7 You want the AI to summarize a staff member's performance review, which includes their name and a medical absence. What should you do first?

1. Check whether your approved tool is cleared for this kind of personal data, and if in doubt, remove identifying details or don't paste it.
2. Paste it; enterprise AI tools are always safe for any data.
3. Paste it, but add "Keep this confidential" to the prompt.
4. Paste it, then delete the chat afterwards.

Exit Ticket

Before you leave

Think of one task you do at least once a month. Which of the six building blocks (role, task, context, format, examples, constraints) have you usually left out when asking AI to help with it, and what one sentence will you add next time?

Key terms

• Prompt: everything you type into an AI tool, including instructions, questions, pasted material and follow-ups.
• Prompting (prompt engineering): the skill of writing instructions that get useful output from an AI.
• Role, task, context, format, examples, constraints: the six building blocks of a good prompt.
• Iteration: improving an answer through follow-up messages in the same conversation.
• Zero-shot prompting: asking for a task with instructions only, no examples.
• Few-shot prompting: including a few examples of the output you want so the AI copies the pattern.
• Chain of thought: asking the AI to work through a problem in steps before answering.
• System prompt / custom instructions: standing instructions that apply to every message in a chat or project.
• Structured output: an answer in a fixed shape, such as a table, form or data format.
• Prompt template: a saved, fill-in-the-blanks prompt for a task you do often.