If a model rushes to answer a word problem, asking it to work step by step may help. Chain-of-thought prompting encourages a model to produce intermediate reasoning before its final answer, rather than jumping straight to a conclusion. A simple prompt is “Let’s think step by step.” This has improved performance on some reasoning problems, including math and logic tasks, but it is not a universal fix. section 7.4 section 2.14
A word problem
Asked how many apples someone has after starting with three and buying five more, a prompted model might state the starting amount, add five, and then answer eight. The intermediate steps make the route to the answer explicit; the answer still needs checking. section 7.4
This is a change to the prompt, not a fine-tuning step. It can be useful with models that do not already reason internally, or when an answer is prematurely shortened or oversimplified. Modern reasoning models may instead generate extended chains of thought as part of answering a question, without needing the same nudge. Their internal reasoning should not be confused with steps a model displays to a user. section 2.14 section 7.4 section 2.7
Table 7.3 Common advanced prompting techniques
| Technique |
Description |
Example prompt |
|---|---|---|
| Chain-of-thought |
Generate step-by-step reasoning before answering |
“Let’s think step by step. John has 3 apples . . .” |
| Tree-of-thought |
Explore multiple reasoning paths in parallel |
“List three strategies, evaluate each, then choose the best.” |
| Self-consistency |
Generate multiple reasoning chains and aggregate |
“Generate five solutions, then select the most common answer.” |
| Reflexion |
Critique and revise output iteratively |
“Review your answer for errors and revise if needed.” |
| Verification |
Add a final self-check before finalizing the response |
“Double-check your answer for accuracy before submitting.” |
| Decomposition |
Break the problem into subtasks executed sequentially |
“Step 1: Extract metrics. Step 2: Summarize trends. Step 3: Write summary.” |
| Least-to-most |
Solve easier subproblems first, then build up |
“First, list known variables. Then calculate step by step.” |
| Pyramid priming |
Build knowledge gradually in layers |
“Step 1: Define carbon neutrality. Step 2: List targets. Step 3: Analyze impact.” |
| Generated knowledge |
Surface key facts before answering |
“List key concepts related to quantum entanglement before answering.” |
| Role-based prompting |
Assign a role or persona to guide behavior |
“You are an experienced employment lawyer advising a client.” |
| Prompt chaining |
Connect multiple prompts into a pipeline |
“Extract relevant quotes → Summarize → Generate final answer.” |
| Conversational context |
Summarize prior dialogue to preserve coherence |
“Summarize the student’s understanding before proceeding.” |
Compare a single step-by-step chain with techniques that explore multiple paths, revise an answer, or add a final check. section 7.4
