Advanced Prompt Engineering: Prompt Chains, Structured Outputs and Fewer Hallucinations

In lesson 3 you learned the six-part formula, which handles most everyday tasks. This lesson opens the Intermediate level with advanced prompt engineering: the techniques people who work with AI daily rely on. You'll split work into steps, control the output format, ground answers in your documents and treat prompts as tools that get tested and improved, in any assistant.

1. Chain prompts
2. Structured outputs
3. Ground in documents
4. Critical roles
5. Test and improve

Prompt chains: one big task, several small steps

When you ask for something complex in a single message ("analyze these 40 reviews, find the problems, propose fixes and write an email to the team"), the AI does everything at once and the quality of each part drops. It works better to split the job into a chain: each step has one clear goal, and its output feeds the next.

Step 1 · Extract
"Read these reviews and list every problem mentioned, with the
exact quote. Don't interpret anything yet."

Step 2 · Group
"Group the problems from the list above into categories. For each
category, say how many reviews mention it and sort them from most
to least."

Step 3 · Propose
"For the 3 most frequent categories, propose 2 solutions each, with
an estimated cost (low, medium, high) and timeline."

Step 4 · Communicate
"Using all of the above, write a 200-word email to the operations
team: the main problem, the data and the proposal."

The advantages: you can fix each step before moving on (an extraction error gets caught before it contaminates the conclusions), each instruction stays simple and the result is far more solid.

Structured outputs: answers ready to use

If you're going to copy the result into Excel, a CRM or another tool, ask for exactly the structure you need. The more precise the template, the less editing you'll do.

Tables

Return a table with these exact columns, in this order:
Date | Customer | Product | Issue | Urgency (High/Medium/Low)
One row per message. If a value isn't there, write "N/A".
Don't add any text before or after the table.

JSON (for automations)

When another program will read the result (for example, in an automation or a Python script), ask for JSON and show the shape:

Extract the data from this invoice and reply ONLY with valid JSON,
no explanations, in this shape:
{"vendor": "", "invoice_number": "", "date": "YYYY-MM-DD",
 "subtotal": 0, "sales_tax": 0, "total": 0}
Use null for any field you can't find.

Expected result for an invoice from a Texas packaging supplier with 8.25% sales tax:

{"vendor": "Lone Star Packaging LLC", "invoice_number": "INV-004587",
 "date": "2026-10-14", "subtotal": 1000.00, "sales_tax": 82.50,
 "total": 1082.50}

Before accepting the extraction, check that subtotal + sales_tax = total ($1,000.00 + $82.50 = $1,082.50). If it doesn't add up, either the model misread a number or the invoice itself has an error: either way, check by hand. Validating what the model returns automatically is exactly what you'll do with Python in lesson 10.

Fixed templates

For documents you produce often (minutes, product sheets, site visit reports), paste the blank template and ask the AI to fill it in: every result comes out in the same easy-to-review format.

Grounding answers in your documents

You saw in lesson 2 that models can invent plausible facts. When you work with documents (contracts, reports, policies, minutes), these instructions sharply reduce the risk:

1. Rule 1Limit the source"Answer only with the information in the attached document."
2. Rule 2Allow "I don't know""If the answer isn't in the document, write 'Not in the document.'" Without that way out, the model tends to fill gaps.
3. Rule 3Ask for exact quotes"For each statement, copy in quotation marks the sentence from the document that supports it, with the section or page." That lets you verify in seconds.
4. Rule 4Extract first, then conclude"First copy the relevant clauses; then, based only on them, answer the question."

<document>
[paste the contract here]
</document>

Question: What's the penalty if I cancel before 12 months?

Instructions:
1. Copy word for word any clauses about cancellation or penalties,
   with their numbers.
2. Answer the question using only those clauses.
3. If the document doesn't specify it, say so clearly.

Critical roles: make the AI push back

By default, assistants tend to agree with you and praise whatever you show them. For important decisions, explicitly ask for the opposite:

Expected result from a premortem: a short story with specific causes ("orders arrived cold because the delivery zone was too large," "the per-order delivery cost ate the margin"). Turn each cause into a question for your plan: what delivery radius is workable? What minimum order still leaves a margin? If the causes are generic ("not enough marketing"), ask it to be specific to your business and your city.

Metaprompting: let AI improve your prompts

AI is very good at writing instructions for itself. Two ways to use that:

IMPROVE A PROMPT
I'm going to use this prompt many times for [task]. Improve it so
the results are more accurate and consistent. Explain what you
changed and why. Current prompt:
"""
[your prompt]
"""

CREATE A PROMPT FROM SCRATCH
I want a reusable prompt for [task]. First ask me whatever you
need to know about the goal, the audience, the format and the
mistakes I want to avoid. Then write the prompt with placeholders
marked [LIKE THIS] for the parts that change each time.

Test your prompts like a pro

A prompt you'll use many times (to sort emails, summarize calls or write product descriptions) deserves a real test, the same way you'd test an Excel formula before filling it down a whole column:

1. Collect 10 to 20 real cases, including hard ones: ambiguous, incomplete, misspelled, in another language or off-topic.
2. Define what a good result is before testing: correct category, no invented data, maximum length, tone.
3. Run the prompt on every case and count how many come out right.
4. Change one thing at a time (add an example, clarify a rule) and rerun the same cases. That's how you know whether the change helped.
5. Save the winning version in your prompt library with the date and what it's for.

Here's how a prompt for sorting support emails might evolve when it's always tested on the same 20 cases (these are the rows already filled in on the Prompt tests tab of the workbook):

Each version changes one thing: 65%, then 80%, then 95% accuracy. Without the fixed set of cases, you wouldn't know which change worked.

Before saving a prompt to your library, check:

• I tested it on at least 10 real cases, including ambiguous and incomplete ones.
• I defined what a good result is before testing (not after seeing the answers).
• It doesn't invent data: when information is missing, it writes "N/A" or says so.
• The format comes out identical in every test, ready to paste where I need it.
• I noted the date, what it's for and which parts change each time.

PROMPT 1 · Extraction
Read the reviews between <reviews> tags and return ONLY a table:
# | Sentiment (Positive/Negative/Mixed) | Main topic | Key quote
Allowed topics: Price, Service, Shipping, Quality, Other.
If a review mentions several topics, pick the main one.
<reviews>
[paste here]
</reviews>

PROMPT 2 · Analysis
Using the table above: count the reviews by topic and sentiment,
sort the topics by number of negative reviews and point out the
topic that's most concerning, with 2 quotes that illustrate it.

PROMPT 3 · Writing
Write a 150-word executive summary for the store manager:
1 sentence with the main conclusion, the data behind it and
2 recommended actions. Direct tone, no unnecessary adjectives.

Summary

• Split complex tasks into prompt chains and review each step.
• Ask for exact structures (tables, JSON, templates) when the result goes into another tool, and validate totals.
• With documents: limit the source, allow "not in the document" and require quotes.
• Ask for critiques, premortems and rubric scores so you don't just get flattery.
• Improve your prompts with AI and test them on a fixed set of cases.

Next lesson: AI for Content Creation: Writing, Social Media, Images and Video. You'll apply these techniques to content with a workflow where you bring the strategy and the expertise.