Course Contents
- 5.1 What Is Prompt Engineering — and Why It Pays
- 5.2 The Four-Part Prompt Structure
- 5.3 Zero-Shot vs Few-Shot Techniques
- 5.4 Chain-of-Thought Prompting
- 5.5 Structured Output: JSON, Tables, Lists
- 5.6 System Prompts and Role Design
- 5.7 Prompt Templates and Variables
- 5.8 Iteration Methodology
- 5.9 Common Mistakes: Five Pitfalls
- 5.10 Template Library: 15 Workplace Scenarios
- 5.11 Multi-Turn Strategy and Context Management
- 5.12 Advanced Optimisation
- 5.13 Summary + Capstone Project
5.1 What Is Prompt Engineering — and Why It Pays
Prompt engineering is the skill of instructing AI precisely enough that its output is usable with minimal editing. It is currently the highest-leverage AI skill because the same model produces 40-point or 90-point results depending solely on the instruction quality — and the instruction costs nothing. Unlike other skills, mastery takes weeks, not years: it is clear thinking, formalised.
5.2 The Four-Part Prompt Structure
[ROLE] You are a senior retail marketing consultant with 10 years of Hong Kong experience.
[TASK] Design 3 Instagram post themes for the Christmas season.
[FORMAT] Each theme includes: a headline, post copy (under 80 words), and 5 hashtags.
[CONSTRAINTS] Casual tone; target audience is cat owners aged 25-40; no emoji.
[EXAMPLE] Last October's post that worked: "Didn't expect lights this fast — oh wait, it's my cat."
- I can write prompts using role/task/format/constraints/examples
- I know missing format specs are the main cause of unstable output
- I saved an industry-specific four-part template of my own
5.3 Zero-Shot vs Few-Shot Techniques
Few-shot solves the brand-voice problem
A pet photography studio has a very distinctive Instagram voice (deadpan Cantonese humour). Generic prompts produced copy that was "too corporate." The owner switched to few-shot: pasting three of her best past captions and asking AI to "match this tone and rhythm for a new post." One shot, perfect match — for AI, examples are the most precise specification you can give.
Choosing zero-shot vs few-shot
- Try zero-shot first: state the need directly. If output scores 80+ and you're done.
- If not, add one example: paste an ideal-output sample and retry.
- Still not there, use 2-3 examples: complex tone and format tasks (brand copy, report style) usually need them.
- Archive the final prompt: this becomes your "brand prompt asset" the whole team reuses.
- I follow the order: zero-shot first, few-shot to fix
- I keep at least one archived few-shot template
5.4 Chain-of-Thought Prompting
I'm pricing a pet shower service. First list the information you need from me.
Then reason step by step: cost structure → competitor reference → psychological price points → finally give me 3 pricing options with their risks.
Do not skip steps; write out every step.
- I know why "step by step" works (forcing explicit reasoning)
- I used CoT on a numeric/comparison/decision task
5.5 Structured Output: JSON, Tables, Lists
| Format | Use case | Prompt phrasing |
|---|---|---|
| List | Ideas, to-dos, check items | "Use a numbered list, one line each, under 15 words" |
| Table | Comparisons, specs, quotes | "Use a markdown table with columns: item/unit price/MOQ" |
| JSON | Feeding another system | "Output JSON only, schema: {name, price, tags[]}" |
| External communication | "Output subject + body; body under 120 words, professional but warm" |
Structured output is the bridge from "using AI" to "automating with AI": lists are for humans, JSON is for machines. Once your prompts stably emit a fixed JSON schema, you can plug AI into spreadsheets, Notion, or your own systems — the workflow-automation course covers this in depth.
- I can specify four output formats by use case
- I tried the "AI outputs JSON → paste into spreadsheet" flow once
5.6 System Prompts and Role Design
The system prompt (or an opening setup block) is your "standing configuration" — write once, effective for the whole conversation. The best-value three lines: (1) role + seniority "You are a Hong Kong retail marketing consultant with 10 years' experience"; (2) output norms "Answer in Traditional Chinese, bold key points, default concise"; (3) boundary "If unsure, say so directly — never invent." These three lines lift everyday conversations immediately.
- I wrote my three-line system prompt and use it daily
- I know role assignment changes the AI's vocabulary and angle
5.7 Prompt Templates and Variables
Variable syntax in three minutes
- Freeze the fixed parts: role, format, constraints stay unchanged.
- Convert changeable parts to [variables]: "Write a [platform] post for [product], targeting [audience]."
- Store in a team-shared doc with fill-in instructions per variable.
- Anyone can use it: colleagues fill the blanks and get 80-point content — that's the leverage of templates.
- I templated my three most-used prompts
- Templates are shared with my team
5.8 Iteration Methodology
Treat the first output as a draft, not a verdict. The fastest improvement path is targeted feedback, not prompt rewrites: "section 2 is too generic — give me two concrete examples with Hong Kong context" beats starting over. Keep a change log for prompts that matter; three iterations usually reach publishable quality.
5.9 Common Mistakes: Five Pitfalls
- I avoid all five pitfalls
- I know iterative revision is the fastest improvement path
5.10 Template Library: 15 Workplace Scenarios
| Scenario | Template core (fill the brackets) |
|---|---|
| Product description | "Write a [platform] product description for [product], highlighting [3 features], tone [brand voice], under [length]" |
| Customer complaint reply | "As [company] support, reply to: [complaint]. Acknowledge, apologise, offer [resolution], under 100 words" |
| Meeting minutes | "Summarise this transcript into decisions / actions+owners / open questions" |
| Weekly report | "Turn these bullets into a weekly report: [bullets]. Sections: wins, numbers, risks, next week" |
| Competitor snapshot | "Compare [us] vs [competitor] on [dimensions]; output a table + 3 takeaways" |
The full 15-template library follows this pattern — copy, fill brackets, run. Build your own variants into the shared library as you go.
5.11 Multi-Turn Strategy and Context Management
- One conversation, one topic — mixing topics degrades relevance as context grows.
- Re-anchor long conversations: every ~10 exchanges, restate the key constraints.
- Carry conclusions forward: when starting fresh, paste the accepted summary rather than replaying the whole chat.
5.12 Advanced Optimisation
- Negative instructions: "avoid jargon, no emoji" works better left explicit.
- Ask for options then critique: "give 3 versions; then list weaknesses of each" — the self-critique pass measurably improves the final pick.
- Temperature awareness: creative tasks welcome randomness; data extraction wants determinism (API temperature 0-0.3).
5.13 Summary + Capstone Project
- Collect your team's 10 most-repeated writing tasks
- Write four-part prompts for each; test until 80+ quality
- Publish the shared library with variable instructions
- Review usage after two weeks and refine the top three
You now direct AI with role-task-format-constraint-example structure, few-shot precision, chain-of-thought reasoning and structured output. Next, put those prompts on autopilot: AI Workflow Automation in Practice →
Keep learning
Frequently Asked Questions
Ready to put AI to work in your business?
Buddylala.ai assigns you a dedicated AI account manager — zero technical overhead, integrating ERP/CRM/Shopify.
Free WhatsApp consultation →