These are the 40 prompts worth saving for the Kimi K2 family, targeting Kimi K2.7, the coding-focused agentic flagship Moonshot AI released in June 2026 (with Kimi K2.6 as the April 2026 general flagship). Both are 1-trillion-parameter Mixture-of-Experts models with about 32B active parameters per token, a 256K-token context, and multimodal input. They are grouped by job: work, coding, agent mode, writing, research, business, study, and data. Run them free at kimi.com on the web or app.
Every prompt is written out in full and ready to copy. The text prompts follow the Role + Task + Context + Constraints + Output format shape. Turn on Thinking mode for the hard ones — it catches edge cases and missing requirements before it commits — and reach for Agent Mode (OK Computer) when you want a finished artifact instead of a chat reply. New here? The Kimi coding prompts go deeper, the agent-mode prompts cover long-horizon work, and the Kimi prompt cheat sheet is a handy one-pager. Swap anything in [BRACKETS] for your own details.
Work & productivity
Give Kimi your real context and it handles the structure. Leave it on Instant for quick drafting and formatting; flip on Thinking when a task hides trade-offs or missing requirements you want surfaced first.
1. Prioritize an overloaded week
Here's everything on my plate this week: [PASTE LIST]. Act as a sharp chief of staff. Sort it by impact and urgency, tell me the three things that actually move the needle, what to delegate or drop, and lay it out as a simple day-by-day plan. Call out anything that looks like busywork.Why it works: The chief-of-staff framing pushes Kimi to cut, not just reorder, an overwhelming list.
2. Meeting notes to action items
Turn these meeting notes into a clean summary: [PASTE NOTES]. Output three sections: Decisions made, Action items (as a table with Owner, Task, Due date), and Open questions. Keep it tight, don't invent details that aren't in the notes, and flag anything that was left unresolved.Why it works: The owner/task/due table turns a wall of notes into something the team can actually track.
3. Draft a difficult work message
Help me write a message to [WHO] about [SITUATION]. My goal is [GOAL] and I want to sound [TONE — e.g. firm but kind]. Keep it honest and clear, don't over-apologize or ramble, and leave the door open for a good response. Give me two versions: a shorter and a slightly warmer one. Context they'll need: [BACKGROUND].Best for: Escalations, pushback, and delicate updates where tone decides how it lands.
4. Summarize a long document
I've attached a document. Read it in full — you have the context window for it — then give me: a 3-sentence summary, the 5 most important points as bullets, any conclusions or recommendations it makes, and anything it leaves unanswered or assumes without evidence. Ground every point in the document only; if something isn't in it, say so rather than filling the gap.Why it works: Kimi's 256K-token context holds long files whole, so it can summarize without dropping the back half.
5. Turn a request into a checklist
Here's a task or request I've been given: [PASTE REQUEST]. Turn it into a clear, ordered checklist of concrete steps. For each step, note what "done" looks like and any input or decision I need from someone else. Flag the steps that are easy to forget, and put anything blocking at the top.Best for: Breaking a vague ask into steps you can actually work through without missing anything.
Coding
Kimi K2.7 is built for coding and long-horizon agentic work. Paste real code, state your language and versions, and give the exact error. Turn on Thinking for gnarly bugs so it surfaces edge cases first. For a deeper set, see the Kimi coding prompts.
6. Debug with a reasoning trace
Turn on Thinking. I'm getting this error in [LANGUAGE / FRAMEWORK version]: [PASTE ERROR]. Here's the relevant code: [PASTE CODE]. Reason step by step through the likely root cause, give me the corrected code, and explain what was wrong in two sentences. If more than one cause is possible, rank them by likelihood. Verify your fix against the error before giving the final answer.Why it works: Thinking mode spends real compute on the trace, which pays off on subtle bugs; the verify step catches confident-but-wrong fixes.
7. Explain unfamiliar code
Explain what this code does to a developer new to [LANGUAGE / LIBRARY]: [PASTE CODE]. Give a one-paragraph overview, then a block-by-block walkthrough of the tricky parts, and call out any bugs, edge cases, or performance concerns you notice. Keep it concrete and skip the obvious lines.Best for: Onboarding to an unfamiliar codebase or reviewing something a teammate wrote.
8. Write tests for a function
Write thorough unit tests for this function using [TEST FRAMEWORK]: [PASTE FUNCTION]. Cover the happy path, edge cases, invalid inputs, and boundary conditions. Explain in one line what each test verifies, and flag any behavior in the function that looks like a bug or is ambiguous enough to need a spec decision.Why it works: Asking Kimi to flag ambiguous behavior surfaces spec gaps while it writes coverage.
9. Refactor without changing behavior
Refactor this code for readability and maintainability without changing behavior: [PASTE CODE]. Improve naming, reduce nesting, and extract logic where it helps. Return the refactored version, then a short bullet list of what you changed and why. Keep it idiomatic for [LANGUAGE] and don't add dependencies.Best for: Cleaning up code you have to keep working in, safely enough to merge.
10. Build a feature from a spec
Turn on Thinking. Build [FEATURE] in [LANGUAGE / STACK]. Requirements: [LIST REQUIREMENTS]. Constraints: [PERFORMANCE, DEPENDENCIES, STYLE]. First outline your approach and the key decisions, then write the full, runnable code with comments on the non-obvious parts. List any assumptions you made and how I'd test it. Verify the code meets each requirement before finishing.Why it works: Outlining the approach first, then coding, catches design mistakes before they're buried in a wall of code.
11. Convert a design image to code
I've attached a screenshot/mockup of a UI. Build it as a responsive [HTML + CSS / React] component that matches the layout, spacing, and typography as closely as you can. Use semantic markup, keep the CSS clean, and note anywhere you had to guess a value. Then list what I'd need to swap in real content or data.Why it works: Kimi is multimodal and strong at design-to-code, so a clear image plus a target stack gets you a working component.
Agent Mode & OK Computer
Agent Mode (OK Computer) is Kimi's agentic workspace. It plans, executes across many steps, and assembles finished Websites, Docs, Slides, and Sheets, running an Agent Swarm of specialized sub-agents. Give it the end goal and constraints, not step-by-step micromanagement. For more, see the Kimi agent-mode prompts.
12. Build a website in Agent Mode
Use Agent Mode. Build a [one-page / multi-page] website for [WHAT — e.g. my freelance service]. Audience: [WHO]. It needs these sections: [LIST]. Tone and style: [DESCRIBE]. Make it responsive, fast, and accessible, write real placeholder copy I can edit, and use a clean modern layout. When it's built, walk me through the structure and tell me exactly what to change to make it mine.Why it works: Agent Mode owns the whole build — layout, copy, styling — so a clear goal plus a section list gets you a real site, not a snippet.
13. Generate a slide deck
Use Agent Mode. Build a [10]-slide deck on [TOPIC] for [AUDIENCE / PURPOSE — e.g. an investor pitch]. Source material: [PASTE OR ATTACH]. One clear idea per slide, a strong title line on each, speaker notes underneath, and a logical arc from problem to ask. Keep the text tight and design it cleanly. Flag any slide where you're missing information from me.Best for: Turning notes or a doc into a presentable deck without building each slide by hand.
14. Ship a working prototype end to end
Use Agent Mode with Thinking. Build a working prototype of [APP / TOOL] in [STACK]. Core features: [LIST THE MUST-HAVES]. Constraints: [DATA, DEPENDENCIES, LOOK]. Plan the build first and show me the plan, then implement it step by step, test as you go, and give me a runnable result plus setup instructions. Note any feature you cut for the prototype and why.Why it works: Long-horizon agentic execution is Kimi's strength; asking for the plan first keeps a multi-step build on track.
15. Run an agent swarm on a project
Use Agent Mode. I want to produce [DELIVERABLE — e.g. a launch kit: landing page, one-pager, and email sequence] for [PROJECT]. Context: [PASTE BRIEF]. Break this into sub-tasks, assign each to a specialized sub-agent, and coordinate them so the pieces are consistent in message and style. Show me the task breakdown, then deliver each part. Keep a single source of truth for the key facts.Best for: Multi-part deliverables where consistency across pieces matters as much as each piece.
16. Automate a repeatable workflow
Use Agent Mode. I do this repetitive task often: [DESCRIBE THE WORKFLOW STEP BY STEP, WITH INPUTS AND OUTPUTS]. Build me a reusable process or small tool that does it with minimal input from me. Show the steps it will take, handle the common edge cases I mention here: [EDGE CASES], and produce the output as [FORMAT]. Tell me what to paste in each time to run it.Why it works: Describing inputs, outputs, and edge cases up front lets the agent build something you can reuse instead of a one-off answer.
Writing
Give Kimi a Role, a clear Task, real Context, and an exact Output format, and it matches your tone closely. Leave Thinking off for straight drafting; turn it on only when the structure itself needs reasoning.
17. Blog post from an outline
You are a writer for [AUDIENCE]. Write a [1,200]-word blog post titled "[TITLE]" from this outline: [PASTE OUTLINE]. Keep the tone practical and direct, lead each section with the takeaway, use short paragraphs and concrete examples, and avoid clichés and filler. Output in clean Markdown with H2 sections and a two-line intro.Best for: Turning a rough outline into a publishable draft that doesn't read like generic AI copy.
18. Rewrite for clarity and tone
Rewrite the text below to be clearer and more direct without losing meaning. Cut filler, fix weak verbs, and keep my key terms. Target a [confident but plain] tone at a [general professional] reading level. Return the rewrite first, then a short bullet list of the main changes you made. Text: [PASTE TEXT]Why it works: Asking for the change list separately keeps the rewrite clean while still showing its work.
19. Cold email in your voice
You are writing a cold outreach email from me ([MY ROLE] at [MY COMPANY]) to [PROSPECT ROLE] at [PROSPECT COMPANY]. Goal: [GOAL]. Keep it under 90 words, open with a specific reason I'm reaching out to them (not a template), state one clear value point, and end with a low-friction ask. Give me three subject-line options.Best for: Outreach that reads like a person wrote it, with subject lines you can A/B test.
20. Headline and hook variations
Generate 12 headline options for [PIECE / PRODUCT] aimed at [AUDIENCE]. Mix angles: benefit-led, curiosity, contrarian, number-led, and how-to. Keep each under 65 characters, avoid hype words, and label the angle of each. Then pick your top three and say why they'd perform best.Why it works: Mixing labeled angles gives you a batch of testable headlines with the reasoning attached.
21. Story or scene draft
Write a [500]-word [genre] scene. Setup: [WHO, WHERE, WHAT'S AT STAKE]. Point of view: [FIRST / THIRD], tense: [PAST / PRESENT]. Show, don't tell; let the tension build through action and dialogue rather than narration; end on a beat that pulls the reader forward. Match this voice: [DESCRIBE OR PASTE A SAMPLE]. Avoid clichés and purple prose.Best for: Fiction drafting when you want a scene that moves instead of a summary of one.
Research & analysis
Turn on Web search for anything current and demand links so every claim is checkable. For a full multi-source write-up, use Deep Research. For documents and data, upload the file and Kimi reads the whole thing first.
22. Cited web research brief
Turn on Web search. Research [TOPIC / QUESTION] and write me a brief as of this month. Cover the current state, the main viewpoints, and where things are heading. Cite every source with a link, prefer primary and recent sources, and note the publication date next to each fact. End with what's still uncertain or contested.Best for: A current, checkable brief instead of an answer from the training cutoff.
23. Deep Research report
Use Deep Research. Produce a thorough report on [QUESTION]. Cover: background, the key data and who's saying what, the main disagreements, and a clear bottom line. Pull from multiple independent sources, cite each with a link and date, and prefer primary sources. Structure it with headings, and end with a "what's still uncertain" section and 3 things I should verify myself.Why it works: Deep Research runs a multi-source, cited investigation, so the structure and citation rules turn it into a report you can actually rely on.
24. Compare sources and find the gap
Turn on Web search. Find the main sources on [QUESTION] and compare what they say. Build a table: Source | Key claim | Evidence given | Date | Link. Then tell me where the sources agree, where they conflict, and which claim is best supported. Flag any source that looks weak or out of date. Cite every source with a link.Why it works: A comparison table across live sources surfaces disagreement that a single blended summary would hide.
25. Fact-check a claim with web search
Turn on Web search. Fact-check this claim: "[CLAIM]". Find the original source, what the primary evidence actually says, and how credible outlets have covered it. Rate it True / Misleading / False / Unproven with a short reason. Cite every source with a link, and separate what's confirmed from what's still uncertain.Best for: Turning a vague "is this true?" into an auditable verdict with links.
26. Analyze an uploaded dataset
I've attached a data file. Read it, then tell me: what the columns mean, the size and shape, the top patterns or trends, any outliers or data-quality issues, and the 3 findings that matter most for [MY GOAL]. Show the numbers behind each finding, and flag anything that would need more data to confirm. Don't invent values that aren't in the file.Why it works: Grounding findings in the file and flagging data-quality issues keeps the analysis honest instead of a confident guess.
Business & strategy
Give Kimi your real context and it handles the structure. Turn on Thinking for strategy calls with real trade-offs; leave it off for straightforward drafting.
27. One-page business plan
You are a pragmatic startup advisor. Draft a one-page plan for this idea: [DESCRIBE IDEA]. Cover: the problem, the target customer, the offer, how it makes money, the first three steps to validate it cheaply, and the biggest risk. Be blunt about weaknesses. Output as clear sections, no filler.Best for: Pressure-testing an idea before you sink time into it. The "be blunt" line matters.
28. Strategy memo with trade-offs
Turn on Thinking. Write a one-page strategy memo on [DECISION / QUESTION]. Context: [BACKGROUND, GOALS, CONSTRAINTS, DATA]. Lay out 2-3 real options, the trade-offs of each, and a clear recommendation with the reasoning. Name the key assumption the recommendation rests on and what would have to be true for it to hold. Keep it tight and executive-ready.Why it works: Thinking mode surfaces the trade-offs honestly instead of arguing for a foregone conclusion.
29. Competitor teardown
Turn on Web search. Analyze [COMPETITOR] for me as of this month. Cover: their positioning, target customer, pricing, main strengths, obvious weaknesses, and where they're vulnerable. Then tell me three specific ways [MY COMPANY / PRODUCT] could differentiate. Cite sources with links, and separate confirmed facts from your inference.Best for: A grounded read on a rival with concrete openings to exploit.
30. Pricing options with reasoning
Turn on Thinking. Help me price [PRODUCT / SERVICE]. Context: [COSTS, TARGET CUSTOMER, COMPETITORS, VALUE DELIVERED]. Propose 2-3 pricing models (e.g. tiers, usage, flat), lay out the trade-offs of each, and recommend one with the reasoning. Include the numbers, name the biggest risk in your recommendation, and suggest one experiment to test it cheaply.Why it works: Multiple models plus a cheap test turns pricing from a guess into a decision you can validate.
31. Investor or client update
Write a monthly [investor / client] update from these notes: [PASTE PROGRESS, METRICS, WINS, PROBLEMS, ASKS]. Structure it: TL;DR, key metrics, what went well, what didn't and what we're doing about it, and specific asks. Keep it honest and skimmable, lead with the numbers that matter, and don't spin the bad news. Target [LENGTH].Best for: A credible update that reports problems straight instead of burying them.
Study & learning
Kimi is a strong tutor because it can reason and adapt. Turn on Thinking when it needs to work a problem alongside you, and upload your course PDF or notes so it teaches from your actual material.
32. Explain a hard concept simply
Explain [CONCEPT] to me. First give the one-sentence core idea, then build it up with a concrete everyday analogy, then a worked example, then the common misconception people have. Assume I know [YOUR CURRENT LEVEL]. End with two questions I could answer to check I actually understood it.Best for: Getting past a concept that a textbook explained badly.
33. Build a study plan
I need to learn [SUBJECT / EXAM] by [DATE]. I can study [HOURS] per week and my current level is [LEVEL]. Build a week-by-week plan: what to cover each week, in what order, and which topics need the most time based on how hard or foundational they are. Include checkpoints to test myself and a lighter final week for review.Why it works: Ordering topics by difficulty and dependency produces a plan that builds, instead of a flat list of chapters.
34. Quiz me with a feedback loop
Quiz me on [TOPIC] at [LEVEL]. Ask one question at a time and wait for my answer before the next. After each answer, tell me if I'm right, correct any mistake with a short explanation, and adjust the difficulty based on how I'm doing. Keep going for [10] questions, then give me a summary of what I've got down and what to review.Best for: Active recall practice that adapts to where you're weak.
35. Socratic tutor session
Turn on Thinking. Be a Socratic tutor for [TOPIC / PROBLEM]. Don't give me the answer. Ask me guiding questions one at a time that lead me to work it out myself, and when I'm stuck, give the smallest hint that unblocks me rather than the solution. If I go wrong, ask a question that helps me see the mistake. Start now.Best for: Learning to solve a type of problem yourself, not just getting this one answer.
Data & spreadsheets
Kimi handles formulas, data cleaning, and analysis, and in Agent Mode it can build a whole Sheet for you. Describe your columns exactly, and upload the file when the data is real.
36. Spreadsheet formula and model
I need a [Google Sheets] formula/model to [GOAL]. My data is laid out like this: [DESCRIBE COLUMNS]. Give me the formula, explain what each part does, and note common ways it could break (blank cells, text vs numbers). If a small helper column would make it cleaner, suggest it.Why it works: The explanation plus gotchas means you can maintain the sheet yourself instead of pasting a formula you don't understand.
37. Build a Sheet in Agent Mode
Use Agent Mode. Build me a spreadsheet to [PURPOSE — e.g. track a budget / plan a project]. Inputs I'll enter: [LIST]. Outputs I want: [LIST — totals, charts, summaries]. Set up the columns, add the formulas, and include a small summary section up top. Explain how to use it in three lines, and tell me which cells I edit versus which are calculated.Why it works: Agent Mode assembles the full Sheet — columns, formulas, summary — so you get a working tool, not a formula to wire up yourself.
38. Clean and reshape messy data
I've attached messy data. Clean it for me: fix inconsistent formats, standardize [DATES / NAMES / UNITS], flag duplicates and blanks, and reshape it into [DESIRED STRUCTURE — e.g. one row per record with these columns]. Return the cleaned table plus a short log of every change you made and anything you couldn't resolve without me deciding.Best for: Getting a messy export into a usable shape without hand-editing every row.
39. Extract structured data from a file
I've attached a file. Extract [WHAT YOU NEED — e.g. every date, name, and amount] and return it as a clean table with columns [COLUMN 1, COLUMN 2, ...]. Read the whole document, don't skip sections, and leave a cell blank rather than guessing if the value isn't stated. After the table, list anything ambiguous that you had to interpret.Why it works: The 256K-token context lets Kimi read the whole file, so extraction doesn't miss data buried deep in a long document.
40. Explain a chart or metric
I've attached a chart/screenshot of a metric. Tell me what it shows in plain English, the trend and any inflection points, what likely explains them, and what I should watch next. Then flag one way this chart could be misleading. If you need context I haven't given (time range, definitions), say what would sharpen the read.Best for: Reading a dashboard or report figure and knowing what actually matters in it.
Want the shortest possible reference for all of this? The Kimi prompt cheat sheet lists every mode and modifier on one page, and bookmark this roundup at /best-kimi-prompts for the full set.
Frequently Asked Questions
Which Kimi model are these prompts built for?
They target the Kimi K2 family, especially Kimi K2.7, the coding-focused agentic flagship from June 2026, with Kimi K2.6 as the April 2026 general flagship. All are 1-trillion-parameter Mixture-of-Experts models with about 32B active parameters per token and a 256K-token context, so the prompts also work on earlier K2 releases.
What is Thinking mode and when should I use it?
Kimi has Instant and Thinking modes. Instant answers fast; Thinking reasons more deeply and catches edge cases and missing requirements before it commits. Turn on Thinking for hard coding, math, planning, and analysis. Leave it on Instant for quick lookups and simple drafting.
What is Agent Mode or OK Computer?
Agent Mode, also called OK Computer, is Kimi's agentic workspace. It autonomously builds Websites, Docs, Slides, and Sheets and runs an Agent Swarm of specialized sub-agents across many steps. Use it for multi-step deliverables where you want Kimi to plan, execute, and assemble a finished artifact rather than a single chat reply.
How do I make Kimi search the web?
Turn on Web search before you send the prompt so Kimi pulls current sources instead of relying on training data. For anything current — prices, news, releases — ask it to cite each source with a link. For a full multi-source report, use Deep Research, which returns a cited write-up across many pages.
Is Kimi free and open-weights?
Yes. Kimi is Moonshot AI's assistant with a free basic tier on the web and app at kimi.com, and the K2 models ship as open weights under a Modified MIT license. Paid plans add higher usage limits. You can start free and run or build on the models yourself.
Can Kimi read my files and images?
Yes. Kimi is multimodal and supports file and image upload. Attach a PDF, doc, spreadsheet, code file, or screenshot and Kimi reads it before you ask. With a 256K-token context it can hold long documents, so you can summarize, extract, compare, or turn a design image into code.
What is the prompt formula for Kimi?
Role + Task + Context + Constraints + Output format. Tell Kimi who it is, what to do, the background it needs, the limits, and the exact shape of the answer. For agentic work, add the end goal and let Kimi plan the steps; for hard problems, turn on Thinking so it surfaces edge cases first.
How do the [BRACKET] placeholders work?
Anything in square brackets is a swap-in. Replace [TOPIC], [PASTE CODE], or [GOAL] with your own details before sending. The rest of each prompt is written to be copy-paste ready.