Plan-and-Solve Prompting
Most reasoning errors happen because models rush into execution without thinking ahead. Plan-and-Solve forces a two-phase approach — first devise a clear strategy, then carry it out step by step — turning impulsive solving into methodical problem resolution.
Introduced: Plan-and-Solve was published in 2023 by Wang et al. The technique was designed to address a key weakness in Zero-Shot Chain-of-Thought prompting: while “Let’s think step by step” encourages sequential reasoning, it provides no guidance on how to structure those steps. Plan-and-Solve replaces that open-ended trigger with a two-phase instruction — “Let’s first understand the problem and devise a plan to solve it. Then, let’s carry out the plan and solve the problem step by step.” The enhanced PS+ variant adds explicit instructions to extract relevant variables, calculate intermediate results, and verify calculations.
Modern LLM Status: The explicit planning paradigm introduced by Plan-and-Solve has been substantially internalized by frontier models. Claude, GPT-4, and Gemini naturally engage in planning behavior when facing complex problems. However, the technique remains valuable when working with smaller models, high-stakes calculations, or problems where you need guaranteed separation between planning and execution phases. The PS+ variant’s emphasis on variable extraction and verification is particularly relevant for mathematical and analytical tasks where precision matters.
Plan First, Solve Second
When you ask a model to “think step by step,” it starts solving immediately — picking an approach on the fly and hoping it leads somewhere useful. This works for simple problems, but for multi-step reasoning, the model often commits to a flawed strategy early and then compounds errors through every subsequent step. The result: confident-sounding answers built on shaky foundations.
Plan-and-Solve separates thinking from doing. Phase one asks the model to understand the problem fully — what information is given, what needs to be found, and what approach makes sense. Phase two then executes that plan methodically. By forcing a pause between strategy and action, the technique catches structural errors before they cascade through the solution.
Think of it like an architect who draws blueprints before construction begins. You would never build a house by laying bricks in whatever direction feels right — you plan the structure first, then execute the plan. Plan-and-Solve applies this same discipline to reasoning.
Zero-Shot CoT’s “think step by step” trigger produces reasoning that is sequential but not necessarily strategic. The model may calculate correctly at each step yet still follow a flawed overall approach. Plan-and-Solve’s explicit planning phase ensures the model commits to a sound strategy before any computation begins, reducing both missing-step errors and calculation mistakes that stem from unclear variable tracking.
The Plan-and-Solve Process
Two distinct phases turning complex problems into structured solutions
Understand the Problem
The model begins by comprehending what is being asked. It identifies the given information, the unknowns, and the constraints. This diagnostic step prevents the model from rushing into calculations before it fully grasps the problem structure.
“A store sells notebooks for $4 each and pens for $1.50 each. Maria buys 3 notebooks and twice as many pens as notebooks. She pays with a $50 bill. How much change does she receive?”
Devise the Plan
Before any computation, the model lays out a step-by-step strategy. This is the signature contribution of Plan-and-Solve — an explicit roadmap that identifies which calculations to perform and in what order. The PS+ variant adds instructions to extract relevant variables and their values at this stage.
Plan: (1) Identify notebook price and quantity. (2) Calculate number of pens from the “twice as many” relationship. (3) Compute total cost of notebooks. (4) Compute total cost of pens. (5) Sum both costs. (6) Subtract from $50 to find change.
Execute the Plan
With the roadmap in hand, the model works through each planned step sequentially. Each calculation follows directly from the plan, ensuring no steps are skipped or performed out of order. Intermediate results are recorded explicitly so they can feed into subsequent steps.
Step 1: Notebooks = 3 at $4 each.
Step 2: Pens = 2 × 3 = 6 pens.
Step 3: Notebook cost = 3 × $4 = $12.
Step 4: Pen cost = 6 × $1.50 = $9.
Step 5: Total = $12 + $9 = $21.
Step 6: Change = $50 − $21 = $29.
Verify and Conclude
The PS+ variant adds a verification phase where the model reviews its calculations for accuracy. This final check catches arithmetic errors and confirms that the answer addresses the original question. The result is a well-grounded conclusion supported by a visible chain of reasoning.
Verification: $12 (notebooks) + $9 (pens) = $21 total. $50 − $21 = $29. The answer checks out. Maria receives $29 in change.
See the Difference
Why explicit planning produces more reliable solutions
Direct Solving
A factory produces 120 widgets per hour. After a 15% efficiency upgrade, it runs for 8 hours but shuts down for 45 minutes due to maintenance. How many widgets were produced that day?
120 widgets × 1.15 = 138 per hour. 8 hours × 138 = 1,104 widgets. Minus 45 minutes downtime... about 1,001 widgets.
Plan-and-Solve
Plan: (1) Calculate upgraded production rate. (2) Convert downtime to hours. (3) Find effective working hours. (4) Multiply rate by effective hours.
Execute:
Upgraded rate: 120 × 1.15 = 138 widgets/hr
Downtime: 45 min = 0.75 hours
Effective hours: 8 − 0.75 = 7.25 hours
Total: 138 × 7.25 = 1,000.5 → 1,000 widgets
The factory produced 1,000 complete widgets that day (138 per hour for 7.25 effective hours, with partial widget rounded down).
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
48 US states now require AI transparency in key areas. Critical thinking remains your strongest tool against misinformation.
Plan-and-Solve in Action
See how explicit planning improves reasoning across different domains
“A train travels from City A to City B at 80 km/h. The return trip is at 120 km/h. If the total round-trip takes 5 hours, what is the distance between the two cities?”
Understanding: We need to find the distance d between two cities, given two different speeds and a total time of 5 hours.
Plan: (1) Set distance as variable d. (2) Express time for each leg using time = distance/speed. (3) Set up equation: time_1 + time_2 = 5. (4) Solve for d.
Execution:
Time A to B: d/80 hours
Time B to A: d/120 hours
Equation: d/80 + d/120 = 5
Common denominator (240): 3d/240 + 2d/240 = 5
5d/240 = 5
d = 240 km
Verification: 240/80 = 3 hours + 240/120 = 2 hours = 5 hours total. Confirmed.
Answer: The distance between City A and City B is 240 kilometers.
“A company’s revenue grew 20% in Q1, declined 10% in Q2, then grew 15% in Q3. If they started the year at $2 million, what was the Q3 ending revenue, and what is the overall growth rate so far?”
Understanding: We have a starting revenue and three sequential percentage changes. We need to apply them in order and calculate both the final figure and the cumulative growth rate.
Plan: (1) Extract starting revenue. (2) Apply Q1 growth of +20%. (3) Apply Q2 decline of −10% to Q1 result. (4) Apply Q3 growth of +15% to Q2 result. (5) Calculate overall growth rate from start to Q3 end.
Execution:
Start: $2,000,000
After Q1: $2,000,000 × 1.20 = $2,400,000
After Q2: $2,400,000 × 0.90 = $2,160,000
After Q3: $2,160,000 × 1.15 = $2,484,000
Overall growth: ($2,484,000 − $2,000,000) / $2,000,000 = 24.2%
Verification: 1.20 × 0.90 × 1.15 = 1.242, which is a 24.2% increase. Confirmed.
Answer: Q3 ending revenue is $2,484,000, representing a 24.2% overall growth for the year so far.
“A team has 4 developers, 2 designers, and 1 project manager. They need to build a feature with a backend API, a frontend UI, and user testing — each task depending on the previous. If each developer can do 1 unit of work per day, each designer 1 unit per day, and tasks require 8, 6, and 3 units respectively, what is the minimum calendar time to complete all three tasks?”
Understanding: Three sequential tasks with different resource requirements. Tasks are dependent (each must finish before the next starts). We need to calculate elapsed time considering parallel work within each task.
Plan: (1) Determine who can work on each task. (2) Calculate days per task based on available workers and units needed. (3) Sum the sequential task durations since they are dependent.
Execution:
Backend API (8 units): 4 developers working in parallel = 8/4 = 2 days
Frontend UI (6 units): 2 designers working in parallel = 6/2 = 3 days
User Testing (3 units): Assuming PM coordinates with 2 designers = 3/2 = 1.5 days, round up to 2 days
Total: 2 + 3 + 2 = 7 calendar days
Answer: The minimum calendar time is 7 days, with the frontend UI phase being the bottleneck at 3 days.
When to Use Plan-and-Solve
Best for multi-step problems that benefit from upfront strategic thinking
Perfect For
Multi-step calculations where variables need careful extraction and tracking — the original use case where Plan-and-Solve showed its biggest gains.
Problems where later steps depend on earlier results — the plan ensures the correct execution order and prevents skipped dependencies.
When you cannot provide examples but still need structured reasoning — Plan-and-Solve delivers decomposition without requiring few-shot demonstrations.
Models that lack strong built-in planning behavior benefit most from explicit plan-then-execute instructions that scaffold their reasoning process.
Skip It When
Questions that require a single lookup or straightforward answer — planning adds unnecessary overhead when the problem needs no decomposition.
Writing stories, brainstorming, or opinion tasks where rigid planning can constrain the organic flow of creative output.
When response speed is critical — the planning phase adds tokens and processing time that may not be justified for time-sensitive use cases.
Use Cases
Where Plan-and-Solve delivers the most value
Financial Modeling
Multi-step calculations involving revenue projections, compound interest, or budget allocations where each variable must be extracted and tracked through sequential operations.
Data Pipeline Design
Planning transformation steps for data processing workflows where the output of each stage feeds into the next, and the overall architecture must be mapped before implementation.
Diagnostic Reasoning
Systematic troubleshooting of technical issues where a plan prevents jumping to conclusions — first gather symptoms, then hypothesize, then test each hypothesis methodically.
Project Estimation
Breaking down complex project timelines into resource allocation, dependency mapping, and milestone calculation where planning prevents overlooked constraints.
Algorithm Development
Designing algorithms by first outlining the approach, identifying edge cases, and establishing the logical structure before writing any pseudocode or implementation.
Academic Problem Solving
Science and engineering homework where students benefit from seeing the model separate understanding from execution, creating a template for structured problem-solving.
Where Plan-and-Solve Fits
Plan-and-Solve bridges unguided reasoning and full decomposition
The Plan-and-Solve+ (PS+) variant extends the basic trigger with more specific instructions: “Extract relevant variables and their corresponding numerals. Calculate intermediate results. Pay attention to calculation and commonsense.” This additional guidance significantly reduces missing-step errors and calculation mistakes, particularly on mathematical reasoning benchmarks where precision matters most.
Related Techniques
Explore complementary reasoning and decomposition techniques
Plan Your Approach
Apply Plan-and-Solve thinking to your own complex problems or build structured prompts with our interactive tools.