Plan-and-solve prompting is a zero-shot reasoning technique that instructs a large language model to first devise a plan for a problem and then execute that plan step by step before giving the final answer. In prompting, it is a direct refinement of zero-shot chain-of-thought: instead of only asking the model to think step by step, it explicitly separates planning from solving.
The method was introduced in 2023 by Wang and colleagues to address typical failures of zero-shot chain-of-thought on arithmetic and commonsense tasks. Its extended variant, often called PS+ prompting, adds detailed instructions about extracting variables and checking calculations. Both variants need no worked examples, which makes them cheap to apply in artificial intelligence systems where hand-written demonstrations are unavailable.
How plan-and-solve prompting structures reasoning
Plan-and-solve prompting works by appending a trigger instruction to the question that asks the model to understand the problem, devise a plan, and then carry out the plan while solving step by step. Because language models generate text left to right, an explicit plan written early in the output conditions every later token, which keeps the reasoning anchored to a defined sequence of subtasks.
The original study used a two-stage pipeline. In the first stage, the model produces the plan and the full reasoning. In the second stage, a short answer-extraction prompt such as “Therefore, the answer (arabic numerals) is” pulls the final value out of the reasoning, which makes automatic scoring reliable.
The table below summarizes the elements of the two published variants.
| Instruction element | PS | PS+ | Purpose |
|---|---|---|---|
| Understand the problem | Yes | Yes | Forces an initial reading of the task |
| Devise a plan | Yes | Yes | Splits the task into subtasks |
| Carry out the plan step by step | Yes | Yes | Executes subtasks in order |
| Extract relevant variables and numerals | No | Yes | Reduces missed information |
| Calculate intermediate results | No | Yes | Reduces skipped steps |
| Pay attention to calculation and commonsense | No | Yes | Reduces arithmetic slips |
The errors it was designed to reduce
Plan-and-solve prompting targets three error types that an error analysis of zero-shot chain-of-thought identified on math word problems. Each instruction element maps to one or more of these failure modes, which explains why the method is more than a rephrased trigger.
| Error type | Typical symptom | Countermeasure |
|---|---|---|
| Missing-step errors | A required intermediate step is skipped | Explicit plan and intermediate results |
| Calculation errors | Correct method, wrong arithmetic | Instruction to check calculation |
| Semantic misunderstanding | The question is misread | Problem understanding and variable extraction |
In the reported experiments, the method reduced calculation and missing-step errors noticeably, while semantic misunderstanding errors declined less. This pattern is plausible: a plan helps a model organize work it already understands, but it cannot compensate for a misread question.
How it compares with related prompting techniques
Plan-and-solve prompting belongs to the family of structured reasoning prompts, and its closest neighbors differ mainly in whether they need examples and whether decomposition happens in one call or several.
| Technique | Examples needed | Model calls | Core idea |
|---|---|---|---|
| Zero-shot chain-of-thought | None | One plus extraction | Generic step-by-step trigger |
| Plan-and-solve prompting | None | One plus extraction | Plan first, then execute |
| Few-shot chain-of-thought | Several worked examples | One | Imitates demonstrated reasoning |
| Least-to-most prompting | Usually a few | Several | Solves subproblems sequentially |
| Decomposed prompting | Per subtask | Several | Routes subtasks to specialized prompts |
| Self-consistency | Optional | Many samples | Majority vote over reasoning paths |
In the original evaluation on benchmarks such as GSM8K, SVAMP and AQuA, PS+ consistently outperformed zero-shot chain-of-thought and came close to few-shot chain-of-thought on several math datasets, despite using no demonstrations. The techniques are also composable: plan-and-solve prompts can be sampled several times and combined with self-consistency for further gains.
Applying the technique in practice
Plan-and-solve prompting is most useful for multi-step problems with a verifiable answer, such as arithmetic word problems, data transformations, scheduling and structured analysis. A practical template places the question first, followed by an instruction to restate the goal, list relevant quantities, write a numbered plan, execute each step and state the final answer in a fixed format.
| Situation | Recommendation | Reason |
|---|---|---|
| Multi-step math or logic task | Use PS+ | Planning reduces skipped steps |
| No worked examples available | Use PS or PS+ | Zero-shot by design |
| Simple factual lookup | Avoid | Adds tokens without benefit |
| Reasoning-tuned model | Test before adopting | Model may plan internally already |
| Strict output parsing required | Add a fixed answer format | Separates reasoning from result |
Modern reasoning-tuned models often produce an internal plan without being asked, so the benefit of an explicit trigger is smaller and less predictable than on the base models used in the original research. Teams should therefore compare it against a plain prompt on a representative evaluation set rather than assume an improvement.
Limitations and common mistakes
Plan-and-solve prompting increases output length and cost, because the plan and intermediate results add tokens to every response. It also inherits a basic weakness of all chain-of-thought methods: a fluent plan does not guarantee a correct plan, and errors in the first step propagate through the rest.
- Treating the plan as verification: the model does not check its plan unless asked; pair it with chain-of-verification or an external checker for critical outputs.
- Ignoring prompt sensitivity: small wording changes in the trigger can shift accuracy, so test several phrasings.
- Skipping answer extraction: without a fixed final format, downstream parsing becomes unreliable.
- Applying it to open-ended writing: for creative or conversational tasks, an enforced plan often makes output rigid without improving quality.
Understanding it simply
When people tackle a tricky problem, they often do better if they sketch a plan before starting. A student solving a word problem might first note what is given, decide which steps are needed, and only then do the arithmetic.
Plan-and-solve prompting asks an AI model to work the same way. The prompt tells it to understand the question, write a short plan, and then follow that plan one step at a time. The main benefit is fewer forgotten steps and fewer careless mistakes, especially in problems with several stages.
Applied in practice
An operations analyst at a logistics company uses a language model to answer questions about delivery capacity, such as how many vans are needed if each van carries 120 parcels, drivers work two shifts, and next week’s forecast is 9,400 parcels per day with a 10 percent safety margin.
With a plain prompt, the model sometimes forgets the safety margin or divides by shifts at the wrong point. The analyst switches to a plan-and-solve template: the model lists the variables, writes a four-step plan, calculates each intermediate value and ends with a line starting with “Final answer:”.
On a test set of 50 similar questions with known answers, the share of correct answers rises clearly, and the remaining errors are easy to locate because each step is visible. The analyst keeps the template for capacity questions and adds a separate calculation check for figures that feed into purchasing decisions.