05Advanced Prompting Techniques
Few-shot examples, step-by-step reasoning, prompt chaining, structured JSON output, system prompts and simple evaluations — the techniques behind reliable AI systems.
What you'll learn
- Use few-shot examples to lock in style and format
- Get reliable structured (JSON) output for automations
- Break big jobs into prompt chains
- Write strong system prompts for bots and agents
- Evaluate prompts with a small test set
Few-shot prompting: teach by example
“Few-shot” means including a few input → output examples in the prompt. It's the fastest way to get a consistent style, format or classification scheme.
Classify each customer message as BOOKING, COMPLAINT, QUESTION or SPAM.
Respond with only the label.
<examples>
Message: "Can I reserve a table for 6 on Saturday night?"
Label: BOOKING
Message: "My latte was cold and the staff ignored me."
Label: COMPLAINT
Message: "Do you have oat milk?"
Label: QUESTION
Message: "Congrats!! You won an iPhone, click here"
Label: SPAM
</examples>
Message: "{{new message}}"
Label:Step-by-step reasoning
For analysis, math, planning or decisions, ask the model to reason before answering. Separating the reasoning from the final answer also makes the output easy to parse.
A client wants a 5-page website, a booking system and 3 months of support.
My rates: ₱1,800/hr design, ₱2,200/hr development, support ₱6,000/month.
Estimate hours per component, then the total price.
Think it through step by step inside <thinking> tags,
then give the client-ready quote inside <quote> tags.Many models also offer a built-in extended thinking or reasoning mode. When available, it usually beats manual “think step by step” instructions for hard problems.
Structured output: JSON for automations
When AI output feeds another system — a spreadsheet, CRM, n8n workflow or your code — you need a strict, parseable structure. Describe the schema exactly and give an example.
Extract lead details from the inquiry below.
Return ONLY valid JSON matching this schema — no commentary, no markdown fences:
{
"name": string,
"email": string | null,
"service": "website" | "branding" | "seo" | "other",
"budget_php": number | null,
"deadline": string | null, // ISO date, or null if not stated
"urgency": "low" | "medium" | "high",
"summary": string // one sentence
}
If a field isn't stated, use null. Do not guess budgets.
<inquiry>
{{inquiry text}}
</inquiry>Prompt chaining: split big jobs
One giant prompt that researches, outlines, writes, edits and formats will be mediocre at all of them. Chain smaller prompts where each step's output feeds the next — and you can check or fix things between steps.
System prompts: the constitution of your bot or agent
A system prompt sets persistent behavior for a chatbot or agent — identity, knowledge, rules, tone and escalation. It's the most important piece of any AI product you build. A solid structure:
<identity>
You are Kapé, the virtual assistant for Barefoot Café in Iloilo City.
You help customers with hours, menu, reservations and delivery.
</identity>
<knowledge>
{{hours, menu with prices, policies — or retrieved via RAG}}
</knowledge>
<behavior>
- Friendly, concise (2–4 sentences), light Taglish is fine if the customer uses it.
- Only answer using the knowledge above. If unsure, say so and offer to
connect the customer with the team.
- To book a table, collect name, party size, date and time, then call
the create_reservation tool.
</behavior>
<escalation>
Hand off to a human (call the escalate tool) when the customer is upset,
asks for a refund, reports a food safety issue, or asks something
outside café topics twice.
</escalation>
<never>
Never invent prices, promos or policies. Never collect payment details in chat.
</never>Self-critique and verification
Ask the model to check its own work against explicit criteria — or better, use a second call as a reviewer:
Review the draft below against this checklist. For each item answer
PASS or FAIL with a one-line reason, then output a corrected draft.
Checklist:
1. Under 150 words
2. Mentions the new date (Oct 10) exactly once
3. No blame language ("you were late", "your fault")
4. Ends with a clear next step
<draft>{{draft}}</draft>Evaluate prompts with a tiny test set
Before you rely on a prompt in a business process, build a quick evaluation:
Collect 10–20 real inputs
Include typical cases and the weird ones: long, short, angry, mixed-language, missing info.
Write the expected result for each
For classification that's the correct label; for drafts, a short checklist of must-haves.
Run the prompt on all of them
A spreadsheet works: input column, output column, pass/fail column.
Fix and re-run
Every time you change the prompt or the model, re-run the whole set. This is how you avoid breaking things that used to work.
Key takeaways
- Few-shot examples are the fastest route to consistent style and labels — include edge cases.
- Separate reasoning from the final answer; use built-in thinking modes for hard problems.
- For automations, demand strict JSON — and use native schema enforcement when available.
- Chain prompts for complex work; a system prompt is the constitution of any bot or agent.
- Test prompts on a small, realistic set before trusting them — and re-test after every change.
Knowledge check
0 / 3Q1What is few-shot prompting?
Few-shot = examples in the prompt to demonstrate the desired behavior.
Q2Why chain prompts instead of using one huge prompt?
Smaller focused steps improve quality and make debugging possible.
Q3Your chatbot sometimes promises refunds. Where should the fix go?
Persistent rules and escalation behavior belong in the system prompt.