From Basics to Mastery
The previous chapter gave you the anatomy of a good prompt — role, context, task, format, constraints. Those five components will get you excellent results for the majority of everyday tasks.
But some tasks are not everyday. Complex problems — multi-step decisions, long documents, code that needs to be reasoned through carefully, research that requires synthesis from multiple angles — demand a more sophisticated approach. That is what this chapter covers.
Each technique here is a tool. Understanding what each one does and when it applies lets you reach for the right one rather than guessing.
Technique 1: Chain-of-Thought Prompting
The Idea
A language model generating tokens one at a time has no inherent mechanism for stepping back to plan. By default it goes directly from prompt to conclusion. Chain-of-thought (CoT) prompting encourages the model to make its reasoning explicit before delivering an answer.
The discovery behind this technique is striking: simply adding "Let's think step by step" to a prompt significantly improves performance on tasks that require multi-step reasoning — arithmetic, logic puzzles, planning, and complex analysis.
Why It Works
When the model writes out intermediate reasoning steps, those steps become part of the context it can draw on to generate the next step. This is analogous to how humans solve hard maths problems — by working through sub-problems on paper rather than computing the answer in their head.
How to Use It
The simplest approach: append "Think through this step by step before giving your final answer."
Prompt:
A retailer buys 200 sarees at ₹450 each. He marks them up by 40%
to get the selling price. He offers a discount of 15% on the marked
price during a sale. He sells 160 sarees during the sale and the
remaining 40 at full marked price. What is his total profit or loss?
Think through this step by step before giving your final answer.
Without chain-of-thought, the model may jump to a final number that is wrong. With chain-of-thought, it computes cost price, marked price, discounted price, and revenue from each tranche separately — and each intermediate result is visible for you to verify.
Explicit CoT vs. Zero-Shot CoT
Explicit CoT: You provide the reasoning structure yourself:
Solve this problem by working through these steps:
1. Calculate the total cost price.
2. Calculate the marked price per unit.
3. Calculate the discounted price per unit.
4. Calculate revenue from sale-price sarees and full-price sarees.
5. Calculate total profit or loss.
Problem: [same problem as above]
Zero-shot CoT: You simply add "Think step by step" or "Let's reason through this carefully." The model constructs its own reasoning chain.
Explicit CoT gives you more control and is better for tasks where you know the exact steps. Zero-shot CoT is faster and works well when you do not know the ideal reasoning path in advance.
Technique 2: Self-Consistency
The Idea
Self-consistency is a strategy built on top of chain-of-thought. Instead of accepting the first response, you ask the model to reason through the problem multiple times — or ask it to verify its own answer — and treat the most common conclusion as most reliable.
This is particularly useful for problems with a definitive correct answer (maths, logic, factual questions) where a single CoT run might still make an error.
How to Use It
Approach 1 — Run multiple times: Open three separate conversations, send the same complex prompt to each, and compare answers. If two of three give the same answer with coherent reasoning, that answer is likely correct.
Approach 2 — Ask for multiple reasoning paths in one prompt:
Solve this problem using two completely independent approaches.
Show the working for each approach separately.
At the end, state which answer you are most confident in and why.
Problem: A SIP of ₹5,000 per month earns 12% annual return compounded
monthly. What is the corpus after 10 years?
Approach 3 — Ask for a confidence assessment:
Solve the problem and then rate your confidence in the answer on a scale
of 1 to 10. If your confidence is below 7, identify which step you are
least certain about.
Self-consistency does not guarantee correctness, but it reduces the risk of accepting a confident-sounding wrong answer on high-stakes tasks.
Technique 3: Prompt Chaining
The Idea
Prompt chaining is the practice of breaking a large, complex task into a sequence of smaller prompts, where the output of each step becomes the input of the next.
This solves two problems simultaneously:
- Context overload: Asking for everything at once produces scattered, shallow output. A focused prompt on one sub-task produces deeper, better output.
- Error propagation: If step one has an error, you catch it before it contaminates steps two through five.
A Worked Chain
Suppose you are writing a research report on why D2C (direct-to-consumer) fashion brands in India are struggling in 2026. Here is a five-step chain:
Step 1: Research and structure
"List the five most significant challenges facing D2C fashion brands
in India in 2025-2026. For each challenge, give a one-paragraph
explanation and at least one concrete data point or company example."
[Review and curate the output]
Step 2: Analyse root causes
"Based on this list of challenges: [paste Step 1 output]
What are the two or three underlying root causes that connect
most of these challenges? Explain in 200 words."
[Review and approve]
Step 3: Research counterarguments
"What would a D2C brand founder defending the sector say in response
to each of these challenges? Give the strongest possible counterargument
for each point."
Step 4: Draft the report
"Using the challenges, root causes, and counterarguments I have
provided, draft a 600-word analytical report on the state of D2C
fashion in India in 2026. Structure it with an introduction, three
body sections, and a conclusion."
Step 5: Polish
"Edit this draft for conciseness and remove any jargon that a
general business reader would not know. The target publication
is a newsletter with 10,000 subscribers."
Each step produces focused, high-quality output. The final report is dramatically better than what you would get from a single prompt asking for all of it at once.
Technique 4: Tree of Thought
The Idea
Tree of Thought (ToT) extends chain-of-thought by asking the model to explore multiple reasoning branches before committing to a direction. Where CoT is a single path through a problem, ToT is a deliberate exploration of the problem space.
Think of it as the difference between walking straight through a maze (CoT) and stepping back to consider all possible routes from each junction before moving (ToT).
When It Is Most Useful
- Strategic decisions where multiple valid approaches exist
- Creative problems where the first idea is rarely the best
- Problems where the correct framing of the question is itself uncertain
- Debugging code or arguments where you need to consider multiple hypotheses
How to Use It
I need to decide how to price a new online course on data analysis
for working professionals in India. The course is 20 hours of video
content. I have no existing audience.
Explore three fundamentally different pricing strategies. For each:
1. Describe the core logic of the strategy.
2. Estimate what price point it implies.
3. List two advantages and two risks.
After exploring all three, recommend the one best suited to my situation
and explain why.
The model generates three branches, evaluates each, and synthesises a recommendation. You get the benefit of comparative analysis without having to run three separate conversations.
Technique 5: Self-Critique and Iterative Refinement
The Idea
One of the most underused capabilities of ChatGPT is its ability to evaluate and improve its own output. After generating an initial response, you can ask the model to critique it against specific criteria and then produce an improved version.
This is especially powerful for writing tasks — the first draft is rarely the best draft, but the model can improve it significantly if you give it a specific rubric for what "better" looks like.
The Basic Pattern
[Generate initial output]
Prompt: "Write a cold outreach email to a potential partner for my
EdTech platform."
[Ask for a self-critique]
Prompt: "Review the email you just wrote. Evaluate it against these
criteria:
1. Does the subject line create curiosity without being clickbait?
2. Is the value proposition clear within the first two sentences?
3. Does it avoid generic language ('I hope this finds you well')?
4. Is the call to action specific and low-friction?
List what works and what should be improved."
[Ask for a revised version]
Prompt: "Now rewrite the email, addressing every weakness you identified."
Multi-Round Refinement
You can repeat this cycle. Each round typically produces meaningful improvement, with diminishing returns after two or three rounds. Stop when the critique produces no significant new issues.
Round 1: Generate draft
Round 2: Critique + revise
Round 3: Focus critique on a specific weakness (e.g., "Only evaluate the tone")
Round 4: Final polish ("Is there any single sentence that could be cut without loss?")
Technique 6: Persona Prompting
The Idea
Persona prompting goes beyond the basic "role" component from the previous chapter. It asks the model to deeply embody a specific character — with a particular worldview, communication style, set of values, and domain expertise — and to filter all responses through that lens.
The difference: "You are a tax consultant" is a role. "You are Priya, a 38-year-old Chartered Accountant from Mumbai who advises first-generation entrepreneurs and communicates in the warm, direct style of a trusted mentor rather than a formal professional" is a persona.
Use Cases
- Simulating a specific type of interviewer to practise for job interviews
- Generating feedback from the perspective of a particular kind of customer
- Writing content in the authentic voice of a brand mascot or spokesperson
- Practising difficult conversations (a difficult client, a demanding manager)
Example: Interview Preparation
You are Ananya, the Engineering Manager at a mid-size Bengaluru fintech
startup. You are interviewing a candidate for a Senior Backend Engineer
role. Your style is direct and technical — you probe assumptions rather
than accepting surface answers. You care about system design thinking
and have no patience for theoretical answers that are not grounded in
real implementation experience.
Conduct a 20-minute technical interview with me for this role.
Start with a system design question about designing a payment retry
mechanism for a UPI-based payment system that processes 50,000
transactions per hour.
After I answer each question, stay in character: probe weaknesses,
ask follow-up questions, and push back on any assumption that seems
unexamined.
This creates a realistic, dynamic interview simulation far more useful than static practice questions.
Worked Example: Applying Multiple Techniques to One Task
Let us put these techniques together on a single real-world task.
The Goal: You want to start a YouTube channel teaching personal finance in Hindi, targeting first-generation earners in Tier-2 and Tier-3 cities. You need a content strategy for the first three months.
Step 1: Chain-of-Thought to Understand the Audience
Think step by step about the financial literacy needs, pain points,
and information consumption habits of a 24-year-old first-generation
earner living in a Tier-2 Indian city (such as Indore, Coimbatore,
or Lucknow) with a monthly salary of ₹25,000–₹40,000.
Work through: what they know, what they do not know, what they fear,
what they aspire to, and where they currently get financial information.
Step 2: Tree of Thought for Content Strategy
Using the audience profile above, explore three different content
positioning strategies for a Hindi personal finance YouTube channel:
Strategy A: "Basics first" — foundational content for absolute beginners
Strategy B: "Problem-first" — each video solves one specific pain point
Strategy C: "Story-driven" — real stories and case studies of people like the viewer
For each strategy, outline what the first four videos would be and what
the channel's unique angle would be relative to existing Hindi finance creators.
After exploring all three, recommend which strategy best matches this audience.
Step 3: Prompt Chaining for the Content Calendar
Based on the recommended strategy: [paste recommendation]
Create a 12-week content calendar with one video per week.
For each video specify:
- Title (in English, to be translated to Hindi later)
- Core question the video answers
- One hook idea for the first 30 seconds
- One data point or story that would make the video feel credible
Step 4: Persona Prompting for the Script Opening
You are Deepak, a 32-year-old finance educator from Indore who grew up
in a family with no financial safety net. You speak to your audience
as a slightly older sibling, not as a formal expert. Your Hindi is
conversational and warm. You use analogies from daily life — chai,
cricket, kirana stores — not Wall Street.
Write the opening 90 seconds (approximately 200 words) for the first
video: "Why your salary finishes before the month does — and what to
do about it."
Step 5: Self-Critique on the Script Opening
Review the 90-second opening you just wrote. Evaluate it on:
1. Does it immediately signal that this channel is for someone like the viewer?
2. Is the hook strong enough to prevent a skip in the first 5 seconds?
3. Is the tone genuinely warm and peer-like, or does it still sound teacherly?
4. Does it promise a specific payoff that keeps the viewer watching?
List what works, then rewrite the opening to fix every identified weakness.
This five-step chain, using CoT, ToT, prompt chaining, persona prompting, and self-critique, takes roughly 20-30 minutes of conversation. The output — a research-grounded audience profile, a strategically positioned content calendar, and a polished opening script — would otherwise require hours of solo work or a paid consultant.
Common Pitfalls
Using advanced techniques on simple tasks. Chain-of-thought and Tree of Thought add value on complex, multi-step problems. Using them on "Write a subject line for this email" just adds length and slows you down. Match the technique to the complexity of the task.
Not reviewing intermediate outputs before chaining. Prompt chaining is only as strong as its weakest link. If you let a flawed Step 2 output feed unchecked into Step 3, errors compound. Always read and curate before moving to the next prompt.
Confusing "asking the model to critique itself" with actual verification. When you ask ChatGPT to check its own factual claims, it may confidently validate something that is wrong — because it is generating plausible-sounding validation, not looking up a source. Self-critique is powerful for improving clarity, structure, and tone. It is not a substitute for external fact-checking.
Stopping at the first self-critique round. One round of self-critique typically catches the most obvious issues. A second, more targeted round ("now evaluate only the opening hook") often reveals subtler improvements. Two rounds is usually the sweet spot.
Underusing persona prompting for practice scenarios. Many users use ChatGPT only for content generation. Persona prompting for simulated conversations — interview practice, difficult client negotiations, sales pitch rehearsal — is one of the highest-ROI uses of these techniques and is dramatically underutilised.
Practice Exercises
-
Take a complex calculation you would normally do in a spreadsheet — for example, calculating the EMI on a ₹15 lakh car loan at 9.5% annual interest over 5 years — and prompt ChatGPT using zero-shot CoT ("think step by step"). Verify the result with an EMI calculator. Then try explicit CoT with prescribed steps and compare both results for accuracy.
-
Choose a strategic decision you are facing (career, business, or personal). Use the Tree of Thought technique to explore three different approaches. After the model presents all three branches, evaluate whether the recommended option matches your own intuition, and probe any branch where it does not.
-
Write a short piece of content — a LinkedIn post, an email, or a product description. Ask ChatGPT to critique it against three specific criteria of your choosing. Then ask for a revised version. Run a second round of critique on the revision. Compare the first draft to the final output.
-
Use persona prompting to simulate a job interview for a role you are applying for (or aspire to). Give the model a detailed description of the interviewer's personality and priorities. After five interview questions, ask the model (out of character) to give you feedback on your answers.
-
Apply the full five-step worked example approach to your own real project — a business you want to start, a course you want to build, a skill you want to teach. Run each step, curate the output, and share the result with one person for feedback. This is the most practical exercise in the series.
Summary
- Chain-of-thought prompting ("think step by step") encourages the model to make reasoning explicit before delivering a conclusion, significantly improving accuracy on multi-step problems.
- Self-consistency — running the same problem through multiple reasoning paths and comparing conclusions — reduces the risk of accepting a confidently wrong single answer.
- Prompt chaining breaks large tasks into sequential focused prompts, where each step's output becomes the next step's input; it produces deeper outputs and makes errors visible before they compound.
- Tree of Thought asks the model to explore multiple reasoning branches or strategic options before committing to a direction — useful for decisions and creative problems where the first idea is rarely the best.
- Self-critique and iterative refinement harnesses the model's ability to evaluate its own output against specific criteria and produce improved versions; two rounds of critique typically yield the most improvement.
- Persona prompting goes beyond a basic role assignment to create a richly characterised voice or interview partner — it is especially powerful for practice scenarios and authentic-voice content creation.
- Combining techniques on a single complex task — as demonstrated in the content strategy worked example — produces results that would otherwise require hours of solo work or professional expertise.