Advanced Prompting | AI Tutorial

Techniques for harder tasks, better reasoning, and expert-level output. Use these when the basics from chapter 02 stop being enough.

Chain-of-Thought (CoT) Prompting

Make the LLM show its reasoning step-by-step. This improves accuracy on multi-step problems, especially math and logic.

Basic CoT

Add the canonical phrase "Let's think step by step" at the end of the prompt. (This wording, from the original Wei et al. paper, is faithfully reproduced because it is the documented trigger phrase.)

Without CoT

What's 15% tip on a $87.50 meal for 3 people per person?

Result: often gets confused.

With CoT

What's 15% tip on a $87.50 meal for 3 people per person?
Let's think step by step.

Result:

15% of $87.50 = $13.13
Total = $100.63
Per person = $33.54

Use it for math, logic, and any multi-step reasoning.

Zero-Shot CoT

Add a trigger phrase at the end:

"Let's think step by step"
"Work through this systematically"
"Break this down"

Few-Shot CoT

Show examples with reasoning included:

Question: If a train travels 60 mph for 2.5 hours, how far does it go?
Reasoning: Distance = Speed x Time = 60 x 2.5 = 150 miles
Answer: 150 miles

Question: A store has a 20% off sale. How much do you pay for a $80 item?
Reasoning: Discount = 80 x 0.20 = $16. Final price = 80 - 16 = $64
Answer: $64

Question: [Your actual question]

Tree of Thought (ToT)

Explore multiple reasoning paths, like a decision tree.

I need to [goal]. Explore 3 different approaches:

Approach 1: [method 1]
Pros: ...
Cons: ...
Likely outcome: ...

Approach 2: [method 2]
Pros: ...
Cons: ...
Likely outcome: ...

Approach 3: [method 3]
Pros: ...
Cons: ...
Likely outcome: ...

Now, which approach would you recommend and why?

Use it for strategic decisions, comparing options, and cases where one path is not obviously best.

Self-Consistency

Generate multiple responses and pick the most consistent answer.

For high-stakes tasks:

Ask the same question 3 to 5 times in fresh conversations
Compare responses
The answer or approach that shows up most often is usually the right one

Use it for high-stakes decisions, complex reasoning, and times when accuracy matters more than speed.

Prompt Chaining

Break a complex task into a sequence of simpler prompts, passing the output of one into the next.

Example: Writing a Blog Post

Prompt 1 (Research)

List 10 key points about [topic] that would interest [audience].

Prompt 2 (Outline)

Based on these key points: [output from Prompt 1]
Create a blog post outline with introduction, 3 main sections, and conclusion.

Prompt 3 (Writing)

Using this outline: [output from Prompt 2]
Write the introduction section. Tone: [tone]. Length: 200 words.

Prompts 4-6: write each main section.

Prompt 7 (Conclusion)

Given this content: [all previous sections]
Write a closing section with a call to action.

Benefits:

Better control at each stage
You can review and adjust between steps
Easier to identify and fix issues

ReAct (Reasoning and Acting)

Combine reasoning with actions (tool use, searches, and so on).

Task: Find the current market cap of the top 3 tech companies and compare them.

Thought: I need to identify the top 3 tech companies first.
Action: List the top 3 tech companies by market cap.
Observation: [Result]

Thought: Now I need current market caps. My knowledge is outdated.
Action: [Search for current data or note the limitation]
Observation: [Result]

Thought: Now I can compare them.
Action: Create comparison.
Result: [Comparison]

Use it for multi-step tasks that require both thinking and information gathering.

Constrained Generation

Force a specific output format with strict constraints.

JSON Output

Extract the following from this text and return ONLY valid JSON:
{
  "name": "",
  "date": "",
  "amount": 0,
  "category": ""
}

Text: "Bought lunch at Chipotle on March 15th for $12.50"

Regex-Like Patterns

Generate 5 email addresses following this pattern:
firstname.lastname@company.com

All lowercase, real names, fortune 500 companies.

Forced Choices

Answer ONLY with: "YES", "NO", or "UNCLEAR"

Question: Based on this contract clause, can we terminate early?
[Contract text]

Role-Based Prompting (Advanced)

Go beyond simple roles. Build detailed personas.

You are Dr. Sarah Chen, a venture capitalist with:
- 15 years in enterprise SaaS investing
- Engineering background (MIT CS)
- Portfolio: $500M AUM
- Known for rigorous technical due diligence
- Direct communication style
- Focus: B2B SaaS, infrastructure

I'm pitching you my [product]. What questions would you ask?

Detailed roles constrain the LLM to think and respond in specific ways. Vague roles produce vague responses.

Meta-Prompting

Ask the LLM to help you write better prompts.

I want to [goal]. Help me write a highly effective prompt for this task.
Consider:
- What context you need
- What format would be best
- What constraints matter
- What examples would help

My initial attempt: [your basic prompt]

Or:

I want to achieve [goal]. Before answering, ask me 5 questions
to better understand what I need, then provide your response.

This forces you to think through requirements before getting output, which usually produces a better prompt than your first instinct.

Negative Prompting

Tell the LLM what NOT to do.

Write a professional email to a client about a project delay.

DO NOT:
- Blame anyone
- Make excuses
- Over-promise on new timeline
- Use corporate jargon

DO:
- Take responsibility
- Provide specific new date
- Offer compensation/solution
- Maintain confident tone

Use it when you have had bad outputs before and know what to avoid.

Perspective-Taking

Ask the LLM to consider multiple viewpoints.

Analyze this business decision from three perspectives:

1. Financial perspective: [focus on numbers, ROI]
2. Customer perspective: [focus on user experience]
3. Team perspective: [focus on execution feasibility]

Then synthesize these into a recommendation.

Use it for complex decisions, avoiding blind spots, and whole-picture analysis.

Socratic Method

Have the LLM ask YOU questions to clarify your thinking.

I'm considering [decision]. Don't give me advice yet.
Instead, ask me 5 probing questions to help me think through this clearly.

After I answer, ask 5 more based on my responses.

This is especially useful when you suspect you have not fully thought through a decision.

Build iteration into the prompt itself.

Task: [What you want]

Your approach:
1. First, provide a draft
2. Then, critique your own draft (what's weak, what's missing)
3. Then, provide an improved version
4. Finally, explain why the second version is better

One prompt, multiple passes, higher quality output.

Constitutional AI Approach

Give the LLM principles to follow.

Constitution for this response:
1. Accuracy over creativity
2. Cite limitations when uncertain
3. Provide balanced views
4. Avoid speculation
5. Use simple language

Now, answer: [Your question]

Use it when you need reliable, balanced, well-bounded responses.

Format-First Prompting

Start with the exact format you want, then ask the model to fill it in.

Create a product comparison:

| Feature | Product A | Product B | Winner |
|---------|-----------|-----------|--------|
| [Fill]  | [Fill]    | [Fill]    | [Fill] |
| [Fill]  | [Fill]    | [Fill]    | [Fill] |
| [Fill]  | [Fill]    | [Fill]    | [Fill] |

Products to compare: [Your products]
Features to evaluate: [Your features]

You get exactly the format you want, with no reformatting work after.

Prompt Templates

Build reusable templates for common tasks.

Code Review Template

Review this [language] code for:
- Bugs and errors
- Performance issues
- Security vulnerabilities
- Best practice violations
- Readability concerns

Code:
```[language]
[Your code]
```text

For each issue found:
1. Severity: [Critical/High/Medium/Low]
2. Location: [Line number or function]
3. Issue: [What's wrong]
4. Fix: [How to fix it]
5. Why: [Explanation]

Analysis Template

Analyze [subject] using this framework:

1. SITUATION
   What's the current state?

2. PROBLEMS
   What are the key issues?

3. ROOT CAUSES
   Why do these problems exist?

4. OPTIONS
   What are 3-5 possible solutions?

5. RECOMMENDATION
   What's the best path forward and why?

6. NEXT STEPS
   What are the immediate actions?

Subject: [Your topic]
Context: [Background info]

Learning Template

Teach me [concept] at [level] using this structure:

1. Simple Definition (one sentence)
2. Why It Matters (real-world relevance)
3. Key Components (break it down)
4. Example 1: [Simple example]
5. Example 2: [More complex example]
6. Common Mistakes (what people get wrong)
7. Quick Quiz (3 questions to test understanding)
8. Next Steps (what to learn next)

Concept: [What you want to learn]
Level: [Beginner/Intermediate/Advanced]

Handling Hallucinations

LLMs confidently generate false information. The next chapter covers this in depth; here are prompt-level mitigations.

1. Request Citations

Explain [topic]. For each claim, cite your source or note if it's
general knowledge vs. uncertain.

2. Ask for Confidence Levels

Answer this question and rate your confidence (1-10) for each part
of your answer.

3. Request Verification Steps

Provide your answer, then list what claims should be verified and how.

4. Use Constraints

Answer only based on: [specific documents you provide]
Do not use any other knowledge.

5. Multiple Models

Check answers across ChatGPT, Claude, and Gemini. Differences indicate
potential hallucination.

Working with Long Context

Modern LLMs have huge context windows (100K to 2M tokens). Use them well.

Structure Long Prompts

# CONTEXT
[Background information - can be pages long]

# TASK
[What you want done]

# SPECIFIC INSTRUCTIONS
[Detailed requirements]

# OUTPUT FORMAT
[How to structure the response]

# CONSTRAINTS
[Boundaries and limitations]

Reference Management

I'll provide several documents. Refer to them as DOC1, DOC2, etc.

DOC1: [Content]
DOC2: [Content]
DOC3: [Content]

Now, compare DOC1 and DOC2 on [criteria], using DOC3 as reference.

Chunking Strategy

For extremely long documents:

I'm going to provide a long document in 5 parts.
After each part, just acknowledge with "Received part X".
After part 5, I'll ask my questions.

Part 1:
[Content]

Multimodal Prompting

Images

[Upload image]

Analyze this image for:
1. Main subject
2. Style/aesthetic
3. Technical quality
4. Potential improvements
5. Suitable use cases

Code and Explanation

[Upload code file]

Review this code and:
1. Explain what it does (high-level)
2. Identify bugs or issues
3. Suggest optimizations
4. Rate code quality (1-10)

Data Analysis

[Upload CSV/data]

Analyze this dataset:
1. Summary statistics
2. Data quality issues
3. Interesting patterns
4. Visualization suggestions
5. Analysis recommendations

Prompt Engineering Patterns

Pattern: Expert Panel

Assemble a panel of 3 experts to evaluate [topic]:
- Expert 1: [Role/specialty]
- Expert 2: [Role/specialty]
- Expert 3: [Role/specialty]

Have each expert provide their perspective, then synthesize into
a final recommendation.

Pattern: Red Team / Blue Team

Analyze this decision:

BLUE TEAM (Advocates for it):
Argue why this is a good idea. Steel-man the position.

RED TEAM (Challenges it):
Argue why this is problematic. Find the weaknesses.

VERDICT:
Balanced assessment considering both perspectives.

Pattern: Time Travel

Project this decision forward:

6 MONTHS FROM NOW:
What likely happened?

1 YEAR FROM NOW:
What are the outcomes?

5 YEARS FROM NOW:
What's the long-term impact?

VERDICT:
Should I do it?

Pattern: Analogical Reasoning

Explain [complex concept] by:
1. Finding 3 analogies from different domains
2. Explaining how each analogy helps understanding
3. Noting where each analogy breaks down

Then synthesize into a clear explanation.

Advanced Control Techniques

Temperature Tuning

0.0-0.2: Facts, code, analysis (deterministic)
0.3-0.5: General writing, instructions (mostly consistent)
0.6-0.8: Creative writing, brainstorming (balanced)
0.9-1.2: Very creative, unconventional ideas (unpredictable)
1.3-2.0: Experimental, chaotic (rarely useful)

Top-P Tuning

0.1: Only most likely tokens (very focused)
0.5: Moderately focused
0.9: Standard (good balance)
0.95: More variety
0.99: Maximum variety

Start with defaults (temperature 0.7, top-p 0.9). Only adjust if you have a specific need.

System Prompts (API Users)

If you are using an API, you can set a system prompt that persists across all messages.

system_prompt = """
You are a technical writing assistant specializing in API documentation.

Guidelines:
- Use clear, concise language
- Always include code examples
- Format code as markdown with language tags
- Note common pitfalls
- Follow our style guide: [link]

Never:
- Use marketing language
- Make assumptions about user skill level
- Provide incomplete examples
"""

You get consistent behavior without repeating instructions in every prompt.

Debugging Prompts

If you are not getting good results:

1. Check Clarity

Is there any ambiguity?
Have you defined all terms?
Are instructions clear?

2. Add Context

What background is missing?
What assumptions are you making?
What does the LLM need to know?

3. Review Format

Have you specified output format?
Is the format realistic?
Have you shown an example?

4. Test Systematically

Original prompt: [X]
Problem: [What went wrong]
Hypothesis: [Why]
Modified prompt: [X with change]
Result: [Better/worse]

5. Simplify

Try removing parts of the prompt
Which parts are necessary?
Are you asking too much at once?

Prompt Libraries

Build your own library of tested, effective prompts.

Organization

prompts/
|-- analysis/
|   |-- data-analysis.txt
|   |-- competitive-analysis.txt
|   |-- swot-analysis.txt
|-- writing/
|   |-- email-templates.txt
|   |-- blog-posts.txt
|   |-- social-media.txt
|-- code/
|   |-- code-review.txt
|   |-- debugging.txt
|   |-- optimization.txt
|-- learning/
    |-- concept-explanation.txt
    |-- tutorial-creation.txt
    |-- quiz-generation.txt

Template Format

PROMPT NAME: [Descriptive name]
CATEGORY: [Category]
USE CASE: [When to use this]
TESTED WITH: [Which models]
SUCCESS RATE: [Your experience]

TEMPLATE:
[The actual prompt with [PLACEHOLDERS]]

EXAMPLE:
[Filled-in example]

NOTES:
[Tips, variations, things to watch for]

A/B Testing Prompts

For important use cases, test variations.

Variation A: [Prompt with approach 1]
Variation B: [Prompt with approach 2]

Test each 5 times:
- Which gives better results?
- Which is more consistent?
- Which is faster/cheaper?

Winner: [A or B] because [reason]

Quick Reference

Key advanced techniques

Chain-of-Thought: "Let's think step by step"
Tree of Thought: explore multiple approaches
Self-Consistency: multiple attempts, pick the best
Prompt Chaining: break complex tasks into steps
ReAct: combine reasoning with actions
Constrained Generation: force specific formats
Meta-Prompting: ask the LLM to improve your prompt
Negative Prompting: specify what to avoid
Constitutional AI: give principles to follow
Format-First: start with the desired output structure

Working principles

Use templates for repeatability
Build a prompt library
Test and iterate systematically
Know when to use advanced techniques and when a simple prompt is enough
Balance complexity against maintainability

Next Steps

Continue to 04-tools-platforms.md for the lay of the land: ChatGPT, Claude, Gemini, the APIs, and specialized tools. Pick one technique from this chapter and apply it to a real task today.