Technical GUIDE

Plan-and-Solve Prompting

Plan-and-Solve (PS) prompting tells a language model to first devise an explicit plan and then carry it out step by step, fixing the failures that plain 'let's think step by step' prompting leaves behind.

2 min readLast updated

Overview

It is a simple prompt tweak that meaningfully boosts multi-step reasoning without any extra training.

Deep Dive

Introduced in a 2023 ACL paper by Lei Wang and colleagues, Plan-and-Solve prompting was a response to a specific weakness in zero-shot chain-of-thought: models often skip steps, miscompute, or misread the question. PS replaces the single instruction 'Let's think step by step' with a two-part directive: '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.' An enhanced variant, PS+, adds reminders to extract relevant variables, calculate intermediate results, and pay attention to numbers. On benchmarks like GSM8K and SVAMP, PS+ closed much of the gap with few-shot chain-of-thought while needing no worked examples in the prompt.

Technical Insight

The mechanism is purely in the prompt: by asking for a plan before execution, PS shifts the model's autoregressive generation so it first produces high-level subgoals, which then condition the detailed reasoning tokens that follow. This separation reduces 'missing-step' and calculation errors. PS+ further steers attention by explicitly naming variables and intermediate quantities, acting as a self-generated scaffold rather than relying on hand-written exemplars.

Strategic Impact

Cost and budget

Architecture decisions drive performance and operating cost for years.

Clearer decisions

Technical education helps teams choose the right stack, not just the newest one.

Quality control

Better engineering choices reduce reliability incidents in production.

The Future of Plan-and-Solve Prompting

Plan-and-solve thinking is now baked into agent frameworks and 'reasoning' models that natively separate planning from execution. Expect planning prompts to merge with tool use, self-verification, and tree-search methods, and to become an internal default behavior in trained-to-reason models rather than a manual prompt. The lasting lesson is that decomposing a task before acting is a cheap, broadly transferable reliability gain.

Real-World Implementation

Solving multi-step grade-school math word problems (GSM8K) where the model first lists the quantities, then computes them in order.

Guiding a coding assistant to outline functions and edge cases before writing any implementation code.

Structuring a customer-support agent to first identify the user's underlying goal, then sequence the resolution steps.

Breaking a complex data-analysis request into 'plan the queries' followed by 'run and combine results' phases.

Risks & Guardrails

Optimizing one benchmark can hide broader system weaknesses.

Infrastructure and maintenance costs are often underestimated.

Security and observability gaps can grow as systems become more complex.

Implementation Roadmap

1

Define latency, quality, and cost targets before implementation.

2

Benchmark under realistic load and data conditions.

3

Instrument monitoring for errors, drift, and user impact.

4

Prepare rollback and incident response paths before scaling.

Keep Exploring

Free newsletter

Keep up with AI in 3 minutes a day

One short email each weekday with the three AI stories that actually matter. Free forever, no ads.

One email each weekday. Unsubscribe in one click. We never sell or share your address.

Test yourself

Take the Plan-and-Solve Prompting quiz

Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.

Start quiz

Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation

Next guide

Least-to-Most Prompting

Frequently asked questions

What is Plan-and-Solve Prompting?

Plan-and-Solve (PS) prompting tells a language model to first devise an explicit plan and then carry it out step by step, fixing the failures that plain 'let's think step by step' prompting leaves behind. It is a simple prompt tweak that meaningfully boosts multi-step reasoning without any extra training.

What core change does Plan-and-Solve prompting make to standard zero-shot chain-of-thought?

PS prompting splits the instruction into devising a plan and then carrying it out step by step, instead of the single 'Let's think step by step' line.

Which specific weaknesses of plain chain-of-thought was Plan-and-Solve designed to address?

The authors targeted missing-step errors and arithmetic/calculation mistakes that zero-shot chain-of-thought frequently produced.

What does the enhanced PS+ variant add to the basic Plan-and-Solve prompt?

PS+ adds detailed reminders to extract relevant variables, calculate intermediate results, and pay attention to the numbers, sharpening the self-generated scaffold.

On which type of benchmark was Plan-and-Solve prominently evaluated?

PS and PS+ were tested on math reasoning benchmarks such as GSM8K and SVAMP, among other reasoning datasets.

Why does generating a plan before the detailed steps tend to help an autoregressive model?

Because generation is autoregressive, the high-level plan tokens become context that guides the subsequent step-by-step reasoning, reducing skipped steps.