Kimi K3.1 Pro · prompting-guide
Kimi recommends stating the task, context, role, input boundaries, execution steps, reference text, examples, and output constraints clearly; complex or long-context tasks should also be categorized, summarized, or processed in chunks to reduce model guesswork.
Kimi recommends stating the task, context, role, input boundaries, execution steps, reference text, examples, and output constraints clearly; complex or long-context tasks should also be categorized, summarized, or processed in chunks to reduce model guesswork.
Suitable tasks: Kimi API requests that require a stable format, clear boundaries, reference material, or multi-step processing; also suitable for designing input workflows for long conversations and long documents.
Unsuitable tasks: Treating general prompting methods as K3.1 Pro-specific capabilities, performance benchmarks, success-rate guarantees, or API parameter-tuning conclusions.
Applicable model version: The Kimi API documentation does not limit this page to a specific model version; the page header includes an announcement for the Kimi K3.1 Pro launch, but the body is a general prompting guide.
Applicable client, agent, or API: Kimi API Chat Completions messages, especially system and user messages; each client still needs to assemble requests according to its own interface.
Recommended reasoning level and parameters: Not specified. The source does not publish recommended values for reasoning_effort, temperature, top-p, or other parameters.
The following are short examples or faithful paraphrases from the official page, not tests exclusive to K3.1 Pro:
Turn vague requests into executable requests. The source changes “How to add numbers in Excel?” into a question about summing each row in an Excel spreadsheet and placing the result in the far-right column named “Total.” It also expands “Work report summary” into a summary of 2023 work records, with highlights listed by month, a total of no more than 500 words, and a year-end summary.
Declare the role in messages. The official example states the assistant's identity, language abilities, and safety boundaries in a system message, then asks in a user message, “My name is Li Lei. What is 1+1?” This shows that the role should be placed in the API request's role messages rather than merely implied by the task body.
Use delimiters to separate inputs. The source wraps two articles in <article>...</article> and also shows a segmented format using Abstract: and Title:. Delimiters make the boundary between the content to process and the instructions clear.
Write the processing order as steps. The official example asks the model to first summarize the text inside triple quotes in one sentence and add the Summary: prefix, then translate that sentence into English and add the Translation: prefix.
Use a small number of examples to express style. The source explains that few-shot examples are usually more efficient than listing every permutation, but the examples themselves are placeholders: Respond in a consistent style and Insert text here. The page does not provide verifiable finished outputs.
When constraining length, prefer paragraph or item counts. The page allows writing “two sentences, 50 words” at the same time, but explicitly warns that word counts are not precise enough; models usually follow a specified number of paragraphs or bullet points more easily.
Provide reference text and a missing-answer response. The official example asks the model to answer only on the basis of the article inside triple quotes; when it cannot find the answer, it should output I can't find the answer.
Break down complex tasks. For customer-service scenarios with many independent rules, the source recommends first determining the query category, then applying only the follow-up rules relevant to that category.
Manage long inputs. For long conversations, summarize earlier turns after the context threshold is reached; for long documents, summarize chapter by chapter and then consolidate the summaries, carrying the summary of preceding material into subsequent chapters when necessary.
This article does not invent a complete prompt. When reusing the methods above, combine instructions for your own task while preserving clear boundaries between input data, reference material, and output requirements.
Write the final task and acceptance criteria clearly first, then add the audience, language, and tone; do not begin by piling up role descriptions.
Separate user content, reference material, and instructions with triple quotes, XML tags, or headings, and explain the purpose of each block.
Write multi-step tasks as ordered steps; when branching is needed, classify first and then list the rules for that category.
For factual questions and answers, add a rule to rely only on the reference text and specify a fixed response when no answer is available.
Use one input/output pair that represents a real task to demonstrate the format; use few-shot examples only when a consistent style is needed.
Prefer paragraph, item, or field counts for length constraints; if a word count is included as well, treat it as an approximate requirement.
Summarize earlier turns when a long conversation reaches the context threshold; split long documents by chapter and progressively consolidate the summaries.
In the Kimi API, put stable roles and global rules in the system message, and put the current data and task in the user message; let the client send them according to the actual interface.
Use real inputs to check missing data, boundary conditions, and formatting; this article did not use an API key and therefore makes no claims about online effects, success rates, or latency.
The official page is titled Best Practices for Prompts. Its introduction explicitly says that the goal is to make Kimi's system/user prompts more reliable and controllable through clear instructions, examples, roles, and output constraints.
The page lists nine categories of recommendations: clear instructions, additional detail, role specification, delimiters, explicit steps, output examples, output length, reference text, and breaking down complex tasks; the latter part adds summaries for long conversations and recursive chunked summaries for long documents.
The Excel summation, 2023 work summary, two <article> elements, abstract and title, triple-quoted summary and translation, and I can't find the answer. are all visible page examples or fixed short phrases; the page provides no corresponding scoring, success-rate, or controlled-experiment data.
The API snippet shown on the page uses a messages array with system/user roles, but the page does not publish model parameters, numerical context thresholds, or general performance guarantees.
This is a general prompting methodology page for the Kimi API, not a kimi-k3.1-pro-specific experiment, model card, or benchmark report. The K3.1 Pro launch notice at the top of the page does not change the general positioning of the body text.
Roles, delimiters, steps, examples, and length constraints are writing recommendations, not control protocols that guarantee the model will always comply; actual behavior still depends on the input, model version, client implementation, and task difficulty.
The page publishes prompting methods and local examples, not a complete prompt for every task; this article does not add any built-in model instructions or performance promises.
The source recommends summaries and recursive chunking, but does not specify a universal context threshold, summary format, compression ratio, or quality metric; the caller must validate these parameters for each task.
The page does not publish recommendations for temperature, top-p, reasoning effort, tool configuration, or API pricing; do not infer parameter trade-offs or K3.1 Pro-specific performance conclusions from this guide.
“Examples are more efficient than covering every permutation” is a statement about prompting methods on the page, not an independent controlled-experiment result provided here; the source examples are also mostly placeholder text.
Open the official Best Practices for Prompts page and verify the page title, sections, and short examples above.
Using an authorized Kimi API client, organize system and user messages according to the source's methods; use your own tasks, examples, reference text, and output constraints.
Test versions with and without delimiters, steps, reference text, and output examples separately. Record the complete input, output format, missing-information handling, and failure cases; do not treat a single result as a model-capability conclusion.
Test summary, chunking, and consolidation workflows separately for long conversations and long documents. Record context length, summary loss, and whether the final task was completed.
This article verifies only the official public documentation. It did not make API requests or measure success rate, latency, cost, or K3.1 Pro-specific effects.
Kimi API Platform (Moonshot AI official documentation) · Source date: Not disclosed · Edited: 2026-10-24
Read the original sourceKimi K3.1 Pro
Run this guide in the environment listed above. Downloading does not transfer the template or establish model availability for your account.
