How To Apply Prompt Engineering Techniques Today

The most useful prompt engineering techniques are zero-shot prompting, few-shot prompting, structured prompting, decomposition, chain-of-thought, tree of thoughts, verification and self-critique, context injection and RAG, and ReAct prompting. This guide explains each one with the example the video demonstrates, plus links to the primary research.

Zero-Shot vs Few-Shot Prompting

Zero-shot prompting asks the model to do a task with no examples, while few-shot prompting supplies a handful of input-output examples so the model copies the pattern. Both are foundational prompt engineering techniques, and the difference is simple: examples or no examples.

Zero-shot works when the task is unambiguous. Classifying the review 'this product is amazing, I absolutely love it' as positive, or translating 'the weather is very pleasant today' into French, needs no demonstration. Modern assistants such as ChatGPT handle sentiment analysis and translation reliably without examples, as the video confirms in its live demo.

Few-shot prompting fits tasks where you need a consistent format or a non-obvious mapping. The video routes customer inquiries to departments with three examples: 'I want to buy this product' goes to Sales, 'my payment failed' goes to Billing, 'the app is crashing' goes to Technical Support. New queries like 'do you have any discount available?' and 'why was I charged 99 for my subscription?' were then routed correctly to Sales and Billing inside Gemini, the assistant built by Google.

The rule of thumb: start with zero-shot, and switch to few-shot when the model guesses your intent instead of following your format.

Structured Prompting With the RTCICO Framework

The RTCICO framework turns a vague request into a complete specification: Role, Task, Context, Instructions, Constraints, and Output format. Vague prompts get vague answers; structured prompts get direct, usable ones.

The video demonstrates the difference with an HR performance review. 'Write a performance review for an employee' makes the assistant ask clarifying questions. The structured version specifies each part in order:

1. Role: 'Act as an experienced HR manager.'
2. Task: 'Write a performance review for an employee.'
3. Context: 'The employee has completed one year in the company and works as a sales executive. Sales target achievement was 110% and they received positive client feedback.'
4. Instructions: 'Highlight achievements, strengths, and one or two areas for improvement; keep the tone professional and constructive.'
5. Constraints: 'Do not use overly negative language. Do not mention information that was not provided.'
6. Output: 'Write the review in three sections: achievements, strengths, and areas of improvement.'

In the demo, Claude, Anthropic AI assistant, produced the three-section review directly with no follow-up questions once the structure was supplied. Constraints are the part most beginners skip, and they are what stop the model from inventing facts or turning the review into a reprimand.

Decomposition and Chain-of-Thought for Complex Decisions

Decomposition prompting breaks one complex problem into named sub-problems the model solves in order. Chain-of-thought prompting, introduced in the 2022 research paper by Wei and colleagues, guides the model through explicit logical steps before the final answer. Together they are the standard cure for generic, hallucinated answers on multi-part questions.

For a healthcare protein product considering a Delhi launch, the video decomposes the decision into six steps: estimate customer demand, identify major competitors, analyze current market and pricing, estimate operational costs, assess revenue opportunities and risks, and only then produce a final recommendation. Fed a bare 'write a business plan', ChatGPT returned generic filler; fed the decomposed prompt, it searched sources, worked each sub-question, and concluded that Delhi is a good test market but that a cautious, small-scale launch makes more sense than heavy upfront investment.

Chain-of-thought prevents the shortcut failure. In the video's cost scenario, a support employee earns 35,000 per month, three employees are needed, and an AI tool costs 60,000 per month while handling 70% of inquiries. Asked bluntly 'should we replace the support team with AI?', a model may compare 60,000 versus 105,000 and answer yes, missing that the remaining 30% of inquiries still require humans. The chain-of-thought prompt asks the model to calculate the current monthly employee cost, then the hybrid human-plus-AI cost, compare both, and evaluate what each option handles, which produces the correct hybrid recommendation in the demo.

Tree of Thoughts and ReAct Prompting

Tree of Thoughts prompting asks the model to explore several independent approaches, compare them, and pick the best, mirroring a boardroom brainstorm. The underlying technique was formalized in the 2023 Tree of Thoughts paper by Yao and colleagues. ReAct prompting, from the 2022 ReAct paper, interleaves reasoning with actions in a loop, so the model thinks, acts, observes the result, and thinks again.

In the video's sales-drop scenario, one prompt asks the same assistant to analyze a 30% sales decline independently as a CEO, a CTO, and a CMO, then compare the three perspectives and combine them into a phased strategy. The output gives each executive's diagnosis and suggestions, then a merged plan, which is exactly what tree-of-thought structuring is for.

The ReAct demo uses a thought-action-observation loop to analyze Nvidia's recent stock performance against competitors over seven days. The model reasons about what to find, takes an action such as searching, observes what came back, and repeats until it can recommend a short-term position. Its final output in the demo classified Nvidia as a 'watch' rather than an aggressive immediate buy. Use ReAct whenever the task requires the model to gather information before it can conclude, and tree-of-thought when one perspective is not enough.

Verification, Self-Critique, Context Injection, and RAG

Four techniques improve answer quality by adding a check, a critique, or real source material: verification prompting, self-critique prompting, context injection prompting, and RAG prompting. Each handles a different failure mode.

Verification prompting tells the model to check its own calculations before answering. The video's test case: a 5,000 rupee product with 20% off, then an additional 10% off the discounted price. The prompt instructs the model to verify each step and confirm the second discount applies to the discounted price before answering, which yields the correct 3,600 final price instead of a rushed figure.

Self-critique prompting asks the model to write a draft, then criticize it as an expert, then rewrite. For a LinkedIn post about learning AI, the demo asks the model to critique its generic first draft as a LinkedIn content expert, identify what is boring or unclear, and rewrite. The revised post is more specific and engaging because the weaknesses were surfaced explicitly.

Context injection means pasting the external facts the model cannot know into the prompt, such as Apple's returns and refunds policy before asking whether an iPhone bought three months ago qualifies for a refund. RAG (retrieval-augmented generation), described in the 2020 paper by Lewis and colleagues, automates that retrieval: give the model your leave-policy PDF, tell it to answer only from the provided context, and it retrieves the sick-leave section and answers from that document, saying 'no information available' rather than guessing.

Rubric-Based Evaluation Prompting

Rubric-based evaluation prompting gives the model a scoring framework so it grades an answer on defined criteria instead of you eyeballing it. It closes the loop: generate, score, improve.

The video asks ChatGPT to write about the economy of India, then prompts: 'Evaluate this answer from one to five for accuracy, clarity, and completeness. Give a reason and an improvement suggestion.' The model returned 4 out of 5 on accuracy, 4 out of 5 on completeness, and higher on clarity, an overall 4.3 out of 5, with concrete reasons for each score.

The practical value is iteration. Once you can see that completeness scored lowest, you can prompt for exactly that gap instead of vaguely asking the model to 'make it better'. Vendor-scored and self-scored, this technique measures its own output, so treat the numbers as a directional signal rather than an independent audit.

Which Technique Should You Use When?

Each technique maps to a specific failure mode, so the choice depends on your task rather than on skill level. The table below summarizes the mapping demonstrated throughout the video.

Technique	Use it when	Demo tool in video
Zero-shot	Task is unambiguous, e.g. sentiment or translation	ChatGPT
Few-shot	You need a consistent format or routing pattern	Gemini
Structured (RTCICO)	Generic answers, missing context or constraints	Claude
Decomposition	Problem has many sub-decisions	ChatGPT
Chain-of-thought	Multi-step math or reasoning risk of shortcuts	ChatGPT
Tree of Thoughts	Multiple stakeholder perspectives needed	ChatGPT
Verification	Calculations must be checked step by step	ChatGPT
Self-critique	Draft output feels generic or weak	ChatGPT
Context injection / RAG	Model needs your private or external documents	ChatGPT
ReAct	Model must reason, act, and observe in a loop	ChatGPT

A useful habit: diagnose why a bad answer failed, then apply the one technique that addresses that failure rather than stacking all techniques into every prompt.

FAQ

• Do I still need prompt engineering when AI can write prompts for me? Yes. The video's argument holds: a calculator did not end the need to understand math, and ChatGPT writing a prompt does not verify the answer's correctness. You need the skills to judge, correct, and improve AI output.

• What is the difference between few-shot and fine-tuning? Few-shot puts examples directly in the prompt at inference time and needs no training. Fine-tuning updates model weights in advance. For most routing and formatting tasks shown here, few-shot prompting is enough.

• Is chain-of-thought prompting the same as decomposition? No. Decomposition splits a large problem into separate sub-problems. Chain-of-thought makes the model reason through explicit logical steps before its final answer. They are often combined on complex decisions.

• What does RAG prompting actually do? Retrieval-augmented generation retrieves relevant passages from documents you provide and uses them as context, so answers come from your source material instead of the model's general knowledge, and it can admit when the context lacks the answer.

Turn Your Own Video Knowledge Into Written Articles

Every technique above arrived as a spoken explanation in a 47-minute video, which is where a lot of good teaching lives and dies. If you have your own recordings, tutorials, or interviews sitting on YouTube, the same gap between what you said and what people can search applies to you.

Skala Blog closes that loop the way chain-of-thought closes a reasoning loop: paste a YouTube URL, get a transcription, and generate a structured, searchable article from it. This draft followed that exact path, and was reviewed like the other Dev doido projects, including the Crazystack typescript blog at crazystack.com.br.

Source video