Skip to main content

Reasoning Strategies

Basic prompts work for simple tasks. For complex problems, you need reasoning strategies. These force the AI to follow specific cognitive patterns before answering.

The difference is noticeable. CoT reduced errors by 40-60% in early GPT-3 tests. ToT solved problems that standard prompting failed on entirely.

tip

For prompt engineering fundamentals (structure, zero-shot vs few-shot, best practices), see Prompt Engineering.

Core reasoning models

StrategyFull NameApproachBest For
CoTChain-of-ThoughtSequential logical stepsMath, planning, logic puzzles
ToTTree-of-ThoughtParallel branching pathsCreative writing, strategic planning
AoTAtom-of-ThoughtDecomposition into atomic partsCoding, proofs, verifiable tasks
CoDChain-of-DraftIterative drafting/refinementLong-form content, essays

Evaluation and refinement strategies

StrategyFull NameApproachBest For
ReflexionReflexionSelf-critique and error correctionCode debugging, quality review
Self-RefineSelf-RefinementIterative self-improvementPolishing drafts, creative work
Self-ConsistentSelf-ConsistencyMultiple attempts + consensusHigh-precision factual tasks

Task-specific strategies

StrategyFull NameApproachBest For
LTMLeast-to-MostSimple to complex sub-problemsEducational tasks, multi-part problems
StandardStandardDirect responseSimple queries, fast summaries

Detailed strategy breakdown

Chain-of-Thought (CoT)

How it works: You tell the model to think out loud, step-by-step.

Why it works: Making intermediate steps explicit cuts hallucinations in logic and math problems.

Prompt pattern:

Let's think through this step by step:
1. First, I'll...
2. Then...
3. Finally...

Best for: Mathematical problems, planning sequences, logical deductions


Tree-of-Thought (ToT)

How it works: The model considers multiple directions, evaluates them, picks the best path.

Why it works: Useful when there's no single correct first step. Get several ideas, then narrow down.

Prompt pattern:

Consider multiple approaches:
- Approach A: [description]
- Approach B: [description]
- Approach C: [description]

Evaluate each and select the best path forward.

Best for: Creative writing, strategic decisions, architectural choices


Atom-of-Thought (AoT)

How it works: Break problems into independent sub-tasks that can be solved separately.

Why it works: Complex software needs interfaces, data structures, and logic defined separately. This forces that separation.

Prompt pattern:

Break this into independent components:
1. [Component A] - can be implemented standalone
2. [Component B] - independent of A
3. [Integration] - how A and B connect

Best for: System design, API design, modular code, mathematical proofs


Chain-of-Draft (CoD)

How it works: Generate a draft, then refine it in multiple passes.

Why it works: Separating generation from refinement produces better long-form content.

Prompt pattern:

Draft 1: Write initial version
Draft 2: Improve clarity and flow
Draft 3: Polish language and style
Final: Review and finalize

Best for: Documentation, essays, complex explanations, technical writing


Reflexion

How it works: Generate a response, critique it, revise based on the critique.

Why it works: First drafts aren't always good enough. This makes the model check its own work.

Prompt pattern:

1. Generate initial solution
2. Critique: What could go wrong? What's missing?
3. Revise based on critique
4. Final answer

Best for: Code review, bug finding, quality assurance, critical analysis


Self-consistent

How it works: Generate multiple attempts at the same problem, find the consensus answer.

Why it works: When you need high confidence in the answer.

Prompt pattern:

Attempt this problem 3 different ways.
Compare the results.
Report the answer that appears most consistently.

Best for: Fact-checking, calculations, verifiable claims


Strategy selection guide

If the task is...Use strategy
"Solve this complex equation"CoT
"Architect a new microservice"AoT
"Give me 5 different story angles"ToT
"Make this code bug-free"Reflexion
"Verify this fact is correct"Self-Consistent
"Write a detailed technical doc"CoD
"Explain this concept step-by-step"LTM
"Quick summary of this file"Standard

Combining strategies

You can stack strategies for complex work:

Code review example:

  1. AoT to break code into reviewable parts
  2. Reflexion to self-critique each part
  3. CoT to walk through the logic

Technical documentation example:

  1. ToT to consider different structures
  2. CoD to draft and refine
  3. Reflexion to check completeness

Using in prompts

Explicit request:

Use Chain-of-Thought reasoning to solve this problem.
Think through each step explicitly before giving the answer.

Implicit cues:

Before answering, consider multiple approaches and evaluate each.
Generate a solution, then critique it for potential issues,
then provide a revised solution.

Summary

NeedStrategyKey Phrase
Step-by-step logicCoT"Think through step by step"
Multiple optionsToT"Consider multiple approaches"
Modular decompositionAoT"Break into independent parts"
Quality iterationCoD"Draft, then refine"
Self-checkingReflexion"Critique your solution"
High accuracySelf-Consistent"Try multiple times, find consensus"