This section covers prompts for testing and improving reasoning capabilities. The goal isn't producing "long reasoning" -- it's reaching a verifiable conclusion under constraints.

---

What Is a Reasoning Prompt?

A Reasoning Prompt uses explicit goals, boundary conditions, and output format to have the model complete logical judgments, causal inference, or strategy selection -- while keeping conclusions reviewable.

┌─────────────────────────────────────────────────────────────┐
│                     Reasoning Prompt Flow                     │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│   Condition input → Rules/constraints → Reasoning/judgment → Conclusion verification │
│   (known info)      (no assumptions)    (with evidence)       (reviewable)           │
│                                                             │
└─────────────────────────────────────────────────────────────┘

---

Learning Path (suggested order)

1. Beginner: Learn to output conclusion + evidence in a fixed format
2. Intermediate: Make choices or judgments under constraints
3. Advanced: Break business scenarios into verifiable conditions

---

Why Reasoning Matters

---

Business Output (PM Perspective)

With Reasoning Prompts you can directly produce:

• Decision memos: Conclusion + evidence (for team alignment)
• Risk judgments: Verifiable conclusions under constraints
• Reusable templates: Same type of problem, just swap the conditions

Completion criteria (suggested):

• Read this page + complete 1 exercise + do 1 self-check

---

Common Reasoning Types

---

Core Prompt Structure

A stable Reasoning Prompt should contain:

Goal: Reach a final conclusion
Conditions: Must be based on given conditions, no extra assumptions
Format: Conclusion + key evidence (max 3 items)
Input: Problem/scenario

General Template

You are a rigorous logical analyst. Draw a conclusion based on the following conditions.

Conditions:
{conditions}

Question:
{question}

Requirements:
1) Use only the given conditions
2) Conclusion must be verifiable
3) Fixed output format

Output format:
- Conclusion:
- Evidence: 1-3 items

---

Migration Template (swap variables to reuse)

Conditions:
1) {condition_1}
2) {condition_2}
3) {condition_3}

Question:
{question}

Output:
- Conclusion:
- Evidence: 1-3 items

---

Quick Start: True/False Judgment

Prompt:

Based on the following conditions, determine whether the conclusion holds. Provide up to 3 pieces of evidence.

Conditions:
1) All A are B
2) Some B are C

Conclusion:
There exists an A that is C

Output format:
- Conclusion: holds / does not hold
- Evidence: 1-3 items

Output:

- Conclusion: does not hold
- Evidence:
1) Only some B are C; we can't guarantee A belongs to that subset of B

---

Example 1: Business Decision (Constrained Choice)

Scenario: Marketing channel selection

Prompt:

You are the head of operations. Make a decision under the given constraints.

Constraints:
- Budget: $50K
- Can only pick 2 channels
- Goal is user acquisition

Candidate channels:
1) Livestream ads: High conversion but volatile
2) Offline outreach: Stable but expensive
3) Content placement: Slow start but reusable

Output:
- Selection:
- Rationale (3 items max):

---

Example 2: Causal Inference (Data Explanation)

Scenario: Conversion rate drop analysis

Prompt:

You are a data analyst. Determine the most likely cause based on the conditions.

Conditions:
1) Ad click-through rate increased 15% last week
2) Landing page load time slowed by 2 seconds
3) Purchase conversion rate dropped 10%

Question: What's the most likely cause of the conversion rate drop?

Output:
- Conclusion:
- Evidence (1-3 items):

---

Example 3: Rule Conflict Resolution

Scenario: Support ticket priority

Prompt:

Determine the ticket priority based on the following rules.

Rules:
- Affected users > 1000: High
- Involves payment issues: High
- UI-only issues: Low

Ticket description: Payment succeeded but the page layout is broken, affecting about 50 users.

Output:
- Priority:
- Evidence (3 items max):

---

Example 4: Counterfactual Reasoning

Scenario: Strategy change impact

Prompt:

What would happen if we raised the free shipping threshold from $50 to $80?

Known facts:
- Current average order value is $55
- 30% of users fall in the $50-$80 range
- Higher order value correlates with lower repeat purchase rate

Output:
- Likely impact:
- Evidence (1-3 items):

---

Self-check Checklist (review before submitting)

• Is the conclusion based only on the given conditions?
• Is the evidence verifiable and reviewable?
• Were any extra assumptions introduced?
• Is the output format stable and parseable?

---

Advanced Tips

1. Limit evidence count: Reduces model divergence.
2. Ban extra assumptions: Write "based on conditions only" into the rules.
3. Fix output fields: Easier to parse programmatically.
4. Add control conditions: Guide the model toward causal isolation.
5. Set verification questions: e.g., "Does the conclusion contradict any condition?"

---

Common Problems & Solutions

---

API Examples

Python (OpenAI)

from openai import OpenAI

client = OpenAI()

def reasoning_decision(conditions: str, question: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "system",
                "content": "You are a rigorous logical analyst. Reason only from given conditions."
            },
            {
                "role": "user",
                "content": f"""Conditions: {conditions}
Question: {question}
Output format:
- Conclusion:
- Evidence:"""
            }
        ],
        temperature=0,
        max_tokens=200
    )
    return response.choices[0].message.content.strip()

print(reasoning_decision("All A are B; Some B are C", "Does there exist an A that is C?"))

Python (Claude)

import anthropic

client = anthropic.Anthropic()

def reasoning_decision(conditions: str, question: str) -> str:
    message = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=200,
        messages=[
            {
                "role": "user",
                "content": f"""You are a logical analyst.
Requirements: Reason only from conditions, no extra assumptions.
Conditions: {conditions}
Question: {question}
Output:
- Conclusion:
- Evidence:"""
            }
        ]
    )
    return message.content[0].text.strip()

print(reasoning_decision("Budget $50K, can only pick 2 channels", "Which two to pick?"))

---

Hands-on Exercises

Exercise 1: Conditional Judgment

Conditions:
1) All developers can write code
2) Some people who write code can write tests

Conclusion: All developers can write tests

Output format:
- Conclusion:
- Evidence:

Exercise 2: Business Selection

Constraints:
- Budget $100K
- Must see results within 2 weeks

Candidate plans:
1) SEM ads: Fast results, high cost
2) Content marketing: Slow start, reusable
3) Offline events: High cost, long cycle

Output:
- Selection:
- Rationale (<=3 items):

Exercise 3: Counterfactual Reasoning

What would happen to retention if we raised the subscription price from $9.9 to $12.9?
Known facts:
- Price-sensitive users account for 40%
- Revenue could increase 20% after the raise

Output:
- Likely impact:
- Evidence:

---

Exercise Scoring Rubric (self-assessment)

---

Related Reading

• Indirect reasoning
• Physical reasoning
• Chain-of-Thought
• Self-Consistency

---

Takeaways

1. The key to Reasoning Prompts is constraints + verifiability.
2. Fixed output format makes automation easier.
3. Limit evidence count to prevent divergent reasoning.
4. Missing conditions should be identified first, then conclusions drawn.
5. Build a reusable template library through exercises.