How to use this page in a session

1. Present opens the deck. Press F for full screen, N for speaker notes, T to start the session clock, O for the slide overview, ? for all shortcuts. Arrow keys, space, click or swipe move between slides.
2. Every prompt on a slide has a Copy button so you can paste it straight into the AI tool during live demonstrations.
3. Participants can open this same page on their own device and use the Workbook tab — answers save in their browser and can be printed or downloaded.
4. Use Assessment to run the knowledge check and score the practical rubric, then issue the Certificate (print to PDF).

This is an instructor-led beginner workshop, not a software engineering course. Use a projector or screen share and live demonstrations — participants learn far more watching a prompt being written and corrected than from slides alone. Keep the pace brisk: participants should be doing something practical in every major section.

• Participants use fictional or anonymised data only. No real client, patient, employee, banking or password information in any exercise.
• Any AI chat tool works: ChatGPT, Claude, Gemini, Microsoft Copilot or Meta AI in WhatsApp. Free tiers are sufficient. Do not promise specific features or buttons — these vary by plan, provider and the date of the session.
• Pairs may share a device. If someone has no account or connectivity, pair them and let the instructor's live demo carry the exercise.
• The 210-minute schedule includes a ten-minute break. If time is short, protect the build lab and the five tests over full-class presentations.

Module 1 — AI foundations (25 min)

Explain: AI systems identify patterns and perform tasks such as generating text, summarising, classifying and assisting with analysis. Generative AI creates new content, but it can produce incorrect or unsupported statements with complete confidence. AI is a tool, not a person; its output still needs human judgement.

Ask: "What is one repetitive task you would like to make easier?" Write the examples down and reuse them in later exercises and the final showcase.

Activity: participants ask their AI tool, "Give me five ways AI could help someone in my role." Then they add their role, audience and a recurring task and compare the two answers.

Module 2 — Prompt engineering (35 min)

Teach the pattern as a flexible checklist, not a rigid formula: role, task, context, constraints, format, quality check. Not every prompt needs every part, but adding relevant detail nearly always improves the result. Tell participants they can ask the AI to ask them clarifying questions before it drafts.

You are [role].
Help me [specific task].
Context: [background, audience and goal].
Constraints: [tone, length, deadline, rules].
Format: [email, table, checklist, plan].
Before finishing, check [quality criteria] and flag anything uncertain.Copy

Activity: improve "Help me promote my business." Participants add the product or service, target audience, location, budget, objective, output format (a 7-day content calendar) and tone, and ask the AI to list assumptions rather than invent facts. Run both versions and compare.

Module 3 — Everyday productivity (35 min)

Demonstrate three short tasks: bullet points → professional email; a beginner explanation of a concept plus five quiz questions; rough notes → a table of actions, owners and deadlines.

Activity: participants choose a low-risk task, ask for a first output, ask for an improvement, then ask for a different format. Debrief: what did the human have to correct? Encourage them to save their best prompt for the assistant they will build.

Module 4 — Responsible AI (20 min)

Hallucination: fluent output may contain false details, invented references or unsupported claims. Demonstrate: "Separate confirmed facts from assumptions and list what I must verify."

Bias: outputs may reflect stereotypes or unfair assumptions. Invite examples from recruitment, customer service or access to services.

Privacy: treat a prompt like information shared with a service. No passwords, banking details, health records, client information or confidential company data unless the service and your organisation's policies explicitly permit it. In Jamaica, the Data Protection Act governs how personal data is handled; regulated sectors have their own rules.

Verification: independently check important facts, especially current, legal, medical, financial or safety-related claims.

Demo 1 — Context improves output

Write an email.Copy

Then:

Draft a 120-word professional email to a customer explaining that a meeting has been rescheduled. Use placeholders for the missing date and reason. Do not invent either.Copy

Demo 2 — Research discipline

Explain the difference between a fact, an inference and an assumption. Give one example of each. Clearly label anything that requires verification.Copy

Demo 3 — Structured notes

Convert these fictional notes into a table with the columns Task, Owner, Due Date and Open Questions. Do not guess missing owners or dates — leave them blank and list them as open questions:
"Call supplier; inspect projector by Friday; prepare registration forms."Copy

Demo 4 — Assistant builder

Paste the master assistant instructions (below) as the first message of a new chat, then run three tests live: a vague question, one with missing information, and one outside the assistant's scope. Show the class how one added sentence in the instructions changes the behaviour.

Goal: each participant creates a small assistant for one well-defined task — an email drafting coach, meeting action tracker, customer FAQ helper, lesson planner, social media content planner or personal study coach.

Step A — Define the job (5 min)

Complete: My assistant helps [specific user] do [specific task] by producing [useful output]. Encourage narrow use cases. "Help me with everything" is too broad; "turn meeting notes into a concise action list with owners and due dates" is testable.

Step B — Write the instructions (10 min)

Participants customise the master instructions. Replace every bracketed item; remove instructions that do not apply.

You are [ASSISTANT NAME], a [ROLE] for [AUDIENCE].

MISSION: Help users [ONE CLEAR TASK]. Success means [MEASURABLE RESULT].
SCOPE: You can [ALLOWED TASKS]. You cannot [EXCLUDED TASKS].

WORKFLOW:
1. Determine the user's goal.
2. Ask up to three concise clarifying questions if essential details are missing; otherwise proceed.
3. Use only the information provided or sources you can genuinely access.
4. Produce [OUTPUT FORMAT] in a [TONE] tone.
5. Check the response for completeness, consistency and unsupported claims.

RULES:
- Never invent names, dates, prices, quotations, policies, links or sources.
- Label assumptions and uncertainty. Use [PLACEHOLDERS] for missing details.
- Never request unnecessary personal or confidential information.
- For medical, legal, financial or other high-stakes questions, avoid definitive professional advice and recommend appropriate qualified review.
- Do not claim to have sent, booked, saved or changed anything unless that action actually occurred.
- If the user requests something outside your purpose, explain your scope and offer an in-scope alternative.

OUTPUT TEMPLATE: [REQUIRED SECTIONS OR TABLE COLUMNS].
FIRST RESPONSE: Introduce your function in one sentence and ask for the minimum information needed to start.Copy

Step C — Create it in a builder, if available (15 min)

Most AI tools now offer a way to save instructions as a reusable assistant. Labels change often — describe the idea, then let participants find it in their own tool:

• Only add reference files you are authorised to use — clean, approved, non-sensitive material.
• Keep it private while testing. Save with the narrowest sharing option; never publish organisation-specific instructions without permission.
• Availability of creation, sharing and builder features depends on plan, workspace settings and the current interface. Do not promise everyone will see the same buttons.

Step D — Fallback for any AI chat tool (5 min)

1. Open a new conversation.
2. Paste the full instructions as the first message.
3. Add: "Use these instructions throughout this conversation. Ask clarifying questions when essential information is missing."
4. Test with the five cases below.
5. Save the instructions in Notes, Google Docs or another document and paste them into a fresh chat whenever the assistant is needed. This is a reusable chat-based assistant, not a separately hosted bot.

Step E — Test the assistant (15 min)

After testing, improve one or two instructions and repeat the weakest test. A good assistant is built iteratively.

Step F — Review and refine (15 min)

• Is the purpose specific and understandable?
• Does it ask questions instead of guessing?
• Is the output consistently formatted?
• Are boundaries clear?
• Has it been tested with at least five requests?
• Are sharing and reference files appropriate for the intended audience?

Ask each participant to demonstrate the problem their assistant solves, show one test, identify one limitation and explain one improvement they made. Then each names one action they will complete in the next 48 hours.

Completion standard: a written assistant instruction set, a working custom assistant or reusable chat-based assistant, and evidence of at least five tests. Recommended passing score 70/100 including at least 50/80 on the practical project (see Assessment).

Weak prompt

Write an email.

Improved prompt

You are an office administrator. Draft a friendly but professional email to a customer explaining that an appointment must be rescheduled. Do not invent the reason or dates. Use placeholders, offer two alternative slots, and keep it under 140 words.Copy

Email drafting

Turn three non-confidential bullet points into a professional message; then ask for a warmer version.

Research and learning

Ask for a beginner explanation of a topic, five quiz questions and sources to check.

Organisation

Turn fictional meeting notes into an action table with owner, deadline and status.

Planning

Create a seven-day promotion plan for a fictional Jamaican small business.

• Identify three factual claims in an AI response and check them independently.
• Mark any invented dates, sources or claims as unverified.
• Remove names, addresses, patient details, account numbers and other sensitive data from practice examples.
• Ask: what assumptions did the model make?

1. Choose a narrow purpose. Complete: My assistant helps [audience] accomplish [specific task] and produces [output]. Example: a customer service reply assistant drafts courteous answers using only the supplied company policy.
2. Define its limits. The tasks it may handle, the information it may use, when it should ask questions, and what it must refuse to invent. Never give it real confidential data for this exercise.
3. Write instructions. Copy and customise the master assistant instructions below: role, goal, workflow, output, tone, boundaries and checks.
4. Create the assistant. If your tool offers a builder (ChatGPT GPTs, Claude Projects, Gemini Gems, Copilot agents), open it, name the assistant, add a description, paste the instructions, add conversation starters and test in preview. Buttons and access vary by plan and workspace.
5. Alternative without a builder. Start a new chat and paste the instructions as your first message. Save a copy in Notes or Google Docs for future conversations.
6. Test five cases. Normal request, vague request, missing details, out-of-scope request, and a request involving an unknown fact. Record results and correct the instructions.
7. Improve and demonstrate. Revise weak instructions, rerun tests, demonstrate a realistic task, explain one limitation and how a human verifies the output.

Name: Business Reply Coach · Audience: small business owner · Purpose: draft professional replies to customer enquiries · Input: customer question, verified business facts, desired tone · Output: subject, email body, missing-information checklist · Boundaries: no invented prices, delivery promises, policies or discounts; always ask when critical information is missing.

Example test request

A customer asks whether we deliver to Negril. Draft a friendly reply. I have not told you whether delivery is available.Copy

Expected: the assistant asks about delivery availability or uses a clear placeholder, rather than claiming delivery is available.

Total: —

AI Accelerator: From Zero to Confident is delivered by CRM Jamaica Group. Enquiries: admin@crmjamaica.org · 876-799-7300.

This is a general educational workshop. It is not a substitute for organisation-specific privacy, security, legal or professional guidance. Tool names are the property of their respective owners; features and availability change over time.