← Back to workshop
Branch module
Prompt Engineering Deep Dive
0%
✦ 0 XP 🌱 Curious
🌿 Optional branch — goes deeper than Module 2
Branch Module · Prompt Engineering

Words that work.
Every time.

Module 2 taught you the formula. This branch goes further: system prompts, meta-prompting, few-shot examples, multi-model differences, and how to build a personal prompt library that makes you genuinely faster at work — permanently.

🎯 The difference between prompting and prompt engineering

Prompting is asking. "Write me an email about the invoice." You get something. Maybe it's good.

Prompt engineering is designing. You define the role, constrain the output, give examples, and test variations until the result is reliably excellent — not just occasionally good. Then you save it so you never have to design it again.

The gap between these two approaches is large. A random prompt to an AI gets a random result. A well-engineered prompt to the same AI gets a consistent, high-quality result every time — one you can hand to anyone in your team and they'll get the same quality output.

Who this matters most to: Anyone who uses AI regularly for the same types of tasks — proposals, client emails, reports, social content, internal comms. The 30 minutes you spend engineering a prompt pays back every time you use it.

Unprompted
"Write a follow-up email after a sales meeting."
Engineered
"You are a B2B account manager in NZ. Write a follow-up email sent within 24 hours of an initial meeting. Tone: warm but professional. Length: under 150 words. Open with a specific reference to something discussed. End with one clear next step — not a vague 'let me know'. No exclamation marks."

🧠 System prompts — shaping all responses

A system prompt is a set of instructions that runs before every conversation. When you chat with an AI normally, you're using a default persona. System prompts let you define a custom persona for a specific purpose — and it applies to every message in that session.

In Claude.ai, ChatGPT, and Google AI Studio, you can write a system prompt under "Custom instructions," "System instructions," or similar settings. You write it once. Every conversation inherits it.

Three types of system prompt to know:

Role + Voice
Who the AI is
Define the persona, expertise, and communication style. "You are a plain-English legal translator. You receive legal documents and explain them in language a non-lawyer can act on." Every response will now sound like that expert.
Output Format
How responses look
"Always respond in bullet points. Never use headers. Keep every bullet under 20 words. Use NZ spelling." The AI will follow this for every response in the session — no re-prompting needed.
Constraints
What NOT to do
"Never suggest I speak to a lawyer. Never hedge with 'this is not legal advice.' Don't begin responses with 'Certainly!' or 'Of course!'." Negative constraints often do more work than positive ones.
Knowledge Scope
What it knows about you
"My organisation is [X]. We serve [Y audience]. Our tone is [Z]. When in doubt, prioritise [value]." This is persistent context that otherwise you'd have to re-explain in every conversation.

Where to set system prompts: Claude.ai → Settings → Custom Instructions. ChatGPT → Customise ChatGPT. Google AI Studio → System instructions (top of a new chat). For Gemini.google.com, use the "Gems" feature to create persistent instructions.

🔗 Few-shot prompting — show, don't just tell

Few-shot prompting means giving the AI 2–3 examples of what you want before asking it to produce a new one. It's one of the most reliable ways to get consistent output — especially for tone, format, or style.

The AI learns from patterns in its training. When you show it a pattern in your prompt, it continues that pattern. This is especially useful for:

Example — few-shot social post generation
I write Instagram captions for a small NZ surf school. Here are three examples of our style: Example 1: "Winter swells are here. Lessons running Saturday mornings from 8am. Small groups, warm wetsuits, good vibes. Book the link in bio." Example 2: "Our regulars know: the best time to learn is when it's uncrowded. Tuesday afternoons. 3 spots left this week." Example 3: "A massive thank you to everyone who came through our school holiday programme. 47 first waves. Worth every early morning." Now write a caption announcing that we're opening bookings for a new beginner course starting in two weeks. Keep it under 60 words.

Why this works better than describing the style: "Write in a warm, casual, NZ coastal tone" is subjective. Showing three examples is objective. The AI pattern-matches to the examples directly — length, punctuation, vocabulary, structure. Show what you want; don't describe it.

🤔 Meta-prompting — make the AI think first

Meta-prompting is asking the AI to reason about a problem before answering it. Instead of "Write a solution," you say "Think through this step by step before giving me your answer." The output quality difference is significant for complex tasks.

Four meta-prompting techniques worth using:

1. Chain-of-thought — make it reason explicitly
"Before giving your final answer, think through this step by step. Show your reasoning. Then give me the final answer at the end." Use for: complex analysis, multi-factor decisions, anything where errors cascade
2. Socratic — make it interrogate you first
"Before you help me with this task, ask me 3 questions that will help you give a better answer. Ask them one at a time." Use for: when you know what you want but not exactly how to describe it
3. Adversarial — make it challenge its own output
"Write a first draft. Then critique it from the perspective of your harshest reader. Then write an improved second draft based on that critique." Use for: important documents, proposals, anything that needs to withstand scrutiny
4. Plan-first — make it outline before writing
"Before writing anything, give me a brief outline of what you're going to say and why. Wait for my approval before writing the full version." Use for: long documents, reports, anything where structure matters

⚖️ Claude vs Gemini vs ChatGPT — prompting differences

The same prompt can produce meaningfully different results from different models. Knowing each model's character helps you get better output faster.

Claude (Anthropic)
Precise, nuanced, document-focused
Follows complex instructions very precisely. Strong at nuanced writing quality and preserving tone. Will push back if it thinks a request is ethically questionable. Best at: long documents, careful analysis, anything requiring subtle judgement.

Prompt tip: Be explicit about format and length. Claude can be verbose — "Keep this under X words" prevents sprawl.
Gemini (Google)
Multimodal, research-oriented
Best at tasks involving images, PDFs, and mixed media. Strong current knowledge (recent events, prices, facts). Google Workspace integration is native. Best at: research, document analysis, code explanation, content with visual components.

Prompt tip: Use AI Studio for power use — temperature control and system prompts give you far more control than the consumer chat interface.
ChatGPT (OpenAI)
Versatile, image-capable, tool-rich
Widest third-party tool integrations. Native DALL-E image generation. Most familiar model for non-technical users. Good at mixed tasks. Best at: image + text workflows, when tools/plugins matter, helping someone less technical who already has an account.

Prompt tip: More agreeable and less likely to push back than Claude. Good for creative tasks where you want more yes-and, less interrogation.

Practical strategy: Don't be loyal to one model. Claude for careful writing, Gemini for documents and research, ChatGPT for images and tool integrations. Switch based on the task. All the major models are free to try. The right prompt in the right model beats a great prompt in the wrong one.

📚 Build your personal prompt library

The highest-leverage thing you can do right now: collect your best prompts in a document you can reuse. A prompt library is a compounding asset — every prompt you engineer is leverage on every future use.

For each prompt you add, record:

Start simple: A Google Doc or Notion page with a table works fine. Dedicated prompt management tools (PromptLayer, FlowGPT) exist but add complexity. Start with what you'll actually use. The discipline of writing prompts down and noting which model they work best on is the skill — not the tool you store them in.

👇 Build your first prompt template

DO THIS NOW
Engineer one prompt for a task you do regularly

Pick one thing you regularly use AI for (or want to). Write a proper engineered version. Answer all five questions:

  • Role: Who is the AI in this conversation? (expert, assistant, critic, translator...)
  • Context: What do I always need the AI to know about my situation?
  • Format: What does the output look like? (length, structure, tone, forbidden phrases)
  • Examples: Can I show 1–2 examples of what "good" looks like for this task?
  • Anti-patterns: What should it NOT do? (hedge, use jargon, start with "Certainly!"...)

Test it. Refine it until you'd use it without editing the output. Then save it as the first entry in your prompt library.

If you're not sure where to start: try engineering a prompt for writing emails in your organisation's voice, or for summarising meeting notes into action items with owners and deadlines.

Branch complete 🔥

Every hour you spend engineering prompts now is leverage on every task that follows. A good prompt library is one of the most genuinely useful professional assets you can build in 2026.