As of March 2026, AI prompting techniques that are good to know

Hello, I'm Lee Sujae from Classmethod.

It seems like generative AI has completely permeated our daily lives, as evident by how easily we can use it on our smartphones that we carry everywhere.
While it's fine to ask questions casually in everyday use, there are situations where high accuracy in responses is necessary, such as when using AI for work.
In such cases, many people use various preliminary processes like RAG, fine-tuning, and prompting.

With AI's rapid advancement, numerous prompting techniques are emerging quickly.
In this article, I'll explore meaningful prompting techniques and those that have become less significant as of March 2026.

What is Prompting?

Prompting is the process of designing and optimizing the text (prompt) input into generative AI models. The quality of results can vary greatly depending on how users question or instruct AI, and systematically researching and applying this is called Prompt Engineering.

AI models generate responses based on learned knowledge and patterns, but even identical questions can yield completely different results depending on how they're expressed.
For example, asking "Find the bugs in this code" will get less useful results than "Analyze this Python code step by step and identify possible errors at each stage with supporting evidence."

Initially viewed as a technique for "finding the right words," recent studies show that the structure and format of prompts and how much context is provided are key. ^[1]

One significant framework change as of 2026 is the clear separation between 'System Instructions' and 'User Prompts'. Constraints, output formats, and persona settings are fixed in system prompts, while user prompts contain only questions and data. For this reason, the concept of "Context Engineering" is gaining attention over "Prompt Engineering."

For more details on Context Engineering, please refer to the article below:
https://news.hada.io/topic?id=21752

Recently Used Prompting Techniques

Here are five effective prompting techniques for improving accuracy, based on papers published between January 2025 and March 2026.

1. Adaptive Graph of Thoughts

Reference Paper: "Adaptive Graph of Thoughts: Test-Time Adaptive Reasoning Unifying Chain, Tree, and Graph Structures" (arXiv:2502.05078, February 2025)

Existing CoT (Chain-of-Thought) and ToT (Tree of Thoughts) approaches had limitations with complex problems due to fixed reasoning structures. AGoT overcomes this by dynamically decomposing problems into sub-problems in the form of a Directed Acyclic Graph (DAG). It works only at test time without additional training and selectively expands only necessary sub-problems to reduce unnecessary computation.

Prompt Example

This is a large-scale migration project from an on-premises environment to AWS for a client targeting a January 2027 opening. Please develop a phased architecture and migration strategy for this project.

Solution approach:

1. Break down the entire migration process into independent subtasks (e.g., DB migration, application containerization, network/security setup, etc.).

2. Specify the dependencies between these subtasks.

3. Sequentially develop solutions for dependent subtasks by referencing the results of preceding tasks.

4. Finally, synthesize a complete migration roadmap.

Additional reference:
https://velog.io/@liger82/paper-Adaptive-Graph-of-Thoughts-Test-Time-Adaptive-Reasoning-Unifying-Chain-Tree-and-Graph-Structures

2. Confidence-Informed Self-Consistency

Reference Paper: "Confidence Improves Self-Consistency in LLMs" (arXiv:2502.06233, ACL 2025 Findings)

Traditional Self-Consistency techniques generate multiple reasoning paths and determine the final answer through majority voting. CISC introduces weighted voting that incorporates model confidence scores for each reasoning path. By giving lower weight to low-confidence answers in voting, it achieves better results with fewer samples.

Prompt Example

Determine whether the following IAM policy adheres to the Least Privilege principle.

Please generate 5 different reasoning paths to answer this question.
For each answer, include a conclusion along with a confidence level between 0 and 100.
Finally, reach a final conclusion by giving more weight to answers with higher confidence levels.

3. Prompt Repetition — Pasting the question twice

Reference Paper: "Prompt Repetition Improves Non-Reasoning LLMs" (arXiv:2512.14982, December 2025, Google Research)

This is the simplest technique to implement. It involves repeating the input prompt exactly twice (<question><question>). Decoder-only LLMs process text sequentially, so when reading the second question, they have already "read" the entire first question, creating a bidirectional context effect.

Prompt Example

What are the optimal solutions for addressing the Cold Start problem in AWS Lambda?
What are the optimal solutions for addressing the Cold Start problem in AWS Lambda?

Additional references:
https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/prompt-iteration?hl=ko
https://news.hada.io/topic?id=26960

4. Adversarial Chain-of-Thought (Adv-CoT)

Reference Paper: "Chain-of-Thought Prompt Optimization via Adversarial Learning" (MDPI Information, December 2025)

This technique automatically improves prompts through adversarial interaction between a generator and discriminator. The generator proposes improvements, the discriminator identifies failure cases, and revisions are made iteratively.

Prompt Example

Please improve the following prompt for better accuracy:
[Current prompt] "Tell me why memory leaks occur in Python code and how to fix them."

Improvement process:

Find three potential failure cases (incomplete aspects) when generating an answer with the above prompt.

Modify the prompt to prevent each failure case.

Generate a response with the modified prompt and explain how it has improved from the original.

5. DR-CoT (Dynamic Recursive Chain of Thought)

Reference Paper: "DR-CoT: dynamic recursive chain of thought with meta reasoning for parameter efficient models" (Scientific Reports / Nature, Vol. 15, 2025)

This technique addresses context dilution and high token costs, which are weaknesses of traditional CoT.

It combines three elements and is designed to perform exceptionally well even on smaller (parameter-efficient) models:

1. Recursive reasoning that breaks problems into sub-problems
2. Dynamic context pruning that maintains only the most important context within a fixed token budget
3. Voting mechanisms that synthesize multiple independent reasoning chains

Prompt Example:

Please solve the problem below, following these rules:

Rules:
- If the problem is complex, break it down into smaller sub-problems.
- When solving each sub-problem, you may reference previous results, but keep only the most essential content and discard unnecessary information (token budget: maximum 150 characters per step).
- Solve the same problem using two different approaches, and if the results match, present that as the final answer.

Problem: A company's annual growth rate was 20% for the first 3 years and -10% for the next 2 years.
If the initial revenue was 10 billion won, what is the revenue after 5 years?

Prompting Techniques No Longer Used

These are prompting techniques that were once effective but have significantly diminished in value or even become counterproductive due to the rapid development of LLMs.

1. Adding "Think step by step" to reasoning models

Reference Paper: "The Decreasing Value of Chain of Thought in Prompting" (arXiv:2506.07142, Wharton Generative AI Labs, June 2025)

Reasoning models like OpenAI o3/o4-mini and Claude Extended Thinking already perform step-by-step reasoning internally. Explicitly instructing CoT to these models is redundant and only increases response time.

Prompt Example
Inefficient method (unnecessary CoT instructions for reasoning models):

Solve this equation: (3x² + 2x - 5) / (x - 1)

Think step by step and explain each solution process. First factor the numerator,
then check if simplification is possible, and derive the final result.

Efficient method (clearly stating only the desired result):

Solve this equation: (3x² + 2x - 5) / (x - 1)
Show the solution process and final result.

2. Role Prompting ("You are an expert in X field")

Reference Paper: "Role-Play Paradox in Large Language Models" (arXiv:2409.13979, updated February 2025)

Role prompting like "You are a cloud architect with 20 years of experience" does not help expand factual accuracy (knowledge boundaries) in current models.
Instead, it risks amplifying biases.

Prompt Example
Less effective approach:

You are a top security expert with 20 years of experience.
Find security vulnerabilities in this AWS IAM policy.

{ "Effect": "Allow", "Action": "*", "Resource": "*" }

Improved approach (providing specific context instead of roles):

Review the following AWS IAM policy from the perspective of the least privilege principle in the AWS Well-Architected Framework.
Tell me about potential security risks and specific improvement measures.

{ "Effect": "Allow", "Action": "*", "Resource": "*" }

3. Excessive Few-Shot Examples (more than 5)

Reference Paper: "The Few-Shot Dilemma: Over-prompting Large Language Models" (arXiv:2509.13196, September 2025)

The "Few-Shot Collapse" phenomenon has been confirmed, where performance drops sharply when the number of examples exceeds a certain level. Latest models already understand tasks, so excessive examples cause overfitting to specific patterns and reduce performance. 2-3 carefully selected examples are sufficient.

Prompt Example:

Excessive Few-Shot (risk of adverse effects):

Classify the sentiment of the following customer reviews (positive/negative/neutral).

Review: "Delivery was fast" → positive
Review: "The packaging was terrible" → negative
Review: "It's just okay" → neutral
Review: "Good quality" → positive
Review: "I want a refund" → negative
Review: "Pretty good for the price" → positive
Review: "I don't think I'll buy it again" → negative
Review: "Not as good as I expected" → negative
Review: "It's average" → neutral
Review: "Highly recommended" → positive

Review: "Not bad at this level" → ?

Appropriate Few-Shot (2-3 examples):

Classify the sentiment of the following customer reviews (positive/negative/neutral).

Review: "Delivery was fast" → positive
Review: "The packaging was terrible" → negative
Review: "It's just okay" → neutral

Review: "Not bad at this level" → ?

4. Complex Prompt Scaffolding for High-Performance Models

Reference Paper: "You Don't Need Prompt Engineering Anymore: The Prompting Inversion" (arXiv:2510.22251, October 2025)

Complex prompts filled with step-by-step rules, constraints, and detailed instruction systems cause a "Prompting Inversion" phenomenon that has adverse effects on top-tier models (GPT-5, Claude Opus level).

Elaborate constraints force "excessive literal interpretation" on high-performance models, hindering autonomous reasoning.
For newer models, it's better to simply and clearly instruct only 'the desired result'.

Prompt Example

Excessively structured prompt (counter-productive for GPT-5 class models):

Follow these instructions in order:
1. First, read the question.
2. Extract keywords related to the question.
3. Define each keyword.
4. Create an outline for your answer based on the definitions.
5. Fill in the outline with complete sentences.
6. Review what you've written and correct any errors.
7. Output the final answer.

Question: What's the difference between REST API and GraphQL?

Simple and clear prompt:

Compare the key differences between REST API and GraphQL from a technical perspective.
Include advantages and disadvantages of each and in which situations one should be chosen.

5. "Magic words" and Emotional Manipulation Phrases

Reference Sources: Wharton GAIL "Prompting Science Report 2" (Meincke, Mollick et al., 2025); Medium "Magic Phrases Don't Work" (January 2026)

Phrases popular in 2023-2024 no longer show consistent effects in current frontier models:

• "Please do this"
• "I'll give you a $200 tip"
• "I'll get fired if I can't do this"
• EmotionPrompt style emotional stimulation phrases ("This is really important for my career")

Initial EmotionPrompt research (Cheng et al., 2023) reported 8-115% improvements in some benchmarks with earlier models, but reproduction experiments with modern models show inconsistent results or minimal effects.
The aforementioned prompt template study (arXiv:2411.10541) supports this by showing that format can make up to 40% performance difference.

Prompt Example

Using emotional manipulation phrases (ineffective)

Please, I'm begging you, help me optimize this SQL query.
I might get fired if I can't fix this at work. This is really important.
Do your absolute best to make it perfect.

SELECT * FROM orders WHERE created_at > '2024-01-01'

Structured and clear approach

Optimize the performance of the following SQL query.

Current issue: Full scan occurs when date filtering the orders table (about 5 million rows)
DB: PostgreSQL 15
Requirement: Reduce response time from 5 seconds to under 500ms

SELECT * FROM orders WHERE created_at > '2024-01-01'

Explain your improvement suggestions along with the expected effects on the execution plan (EXPLAIN).

Conclusion

Writing this article has allowed me to reflect on the prompting methods I've been using.
Looking through various articles, I felt that development is so rapid that if you don't immediately grasp the trends, you might often be using prompts that are not appropriate for the model you're using.

Thank you for reading this long article.

References

For inquiries, contact Classmethod Korea!

Classmethod Korea conducts various seminars and events.
Please refer to the page below for ongoing events.

https://classmethod.kr/board/library

For consultations about AWS and inquiries about Classmethod Members, please contact us at:
Info@classmethod.kr