Kimi K3 is now availableExplore Kimi K3
Kimi K3 Prompts — Featured Kimi K3 Prompt Examples

Kimi K3 Prompts

Reusable prompt templates adapted from source-backed Kimi K3 use cases — browser games, 3D builds, frontend motion, coding agents, integrations, and deep research.

Browse source-attributed templates, copy the one that fits your workflow, and continue to the Kimi K3 model page for production access through EvoLink’s unified API gateway.

Featured Kimi K3 Prompt Examples

25 reusable templates adapted from public, source-attributed cases. Browse the growing collection in the GitHub use-case library.

Browse Kimi K3 Prompt Categories

Single promptIntermediate

Prototype A Voxel Pod-Racing Game

Turns a compact racing concept into a playable voxel prototype with clear room for a second-pass polish loop.

Build a playable browser-based voxel pod-racing prototype inspired by [SCI-FI THEME]. Start with one complete race loop: responsive steering, acceleration and braking, a readable track, collision, checkpoints, a lap timer, and a restart flow. Generate simple voxel-style racers and scenery in code with no paid assets. Run the game, test the controls, and fix blocking issues before delivery. Then list the highest-value second-pass improvements for more racers, a finish-line sequence, effects, and visual detail.

voxelracing
Single promptAdvanced

Build An Interactive Aircraft Carrier In Three.js

Builds an explorable carrier-deck prototype with aircraft operations, camera controls, and a focused interaction loop.

Create an interactive Three.js prototype of a modern aircraft carrier during [TIME OF DAY]. Include a navigable deck, a simplified carrier silhouette, parked aircraft, moving deck crew markers, ocean motion, lighting, and one complete launch-and-recovery sequence. Add orbit and first-person camera modes plus a compact control guide. Keep geometry efficient enough for a normal laptop. Run the scene in a browser, inspect it from multiple angles, and fix camera, clipping, scale, and performance problems before delivery.

threejssimulation
AgenticExpert

Build A 2.5D Paper-World Game With Agent Tools

Coordinates code, generated assets, browser testing, and iteration to produce a cohesive paper-world game prototype.

Build a playable 2.5D paper-world adventure using [ENGINE OR WEB STACK] and the asset tools available to you. The vertical slice must include: one explorable area, a paper-style player character, layered 2D/3D scenery, one enemy encounter, one collectible, dialogue, and a clear completion state. Keep the visual language consistent across generated sprites, textures, and geometry. Work in an autonomous loop: plan, build, launch, play-test, capture screenshots, diagnose the largest visual or gameplay problem, and iterate until the full slice is playable. Record which assets came from which tool.

2.5dagent tools
AgenticAdvanced

Model A Detailed V8 Engine With Blender MCP

Uses Blender MCP to build, inspect, and refine a recognizable mechanical assembly with named components.

Using Blender through the available MCP tools, model a detailed but efficient V8 engine assembly. Include the block, cylinder banks, heads, intake manifold, pulleys, belts, exhaust headers, and a labeled exploded-view option. Use sensible object names and collections. Apply simple physically plausible materials and studio lighting. Inspect the model from front, side, top, and three-quarter views. Correct intersections, floating parts, broken proportions, and unreadable materials. Save the final scene and report any mechanical details that are illustrative rather than engineering-accurate.

blendermcp
AgenticExpert

Build A Multiplayer Voxel Browser Game

Builds a bounded multiplayer voxel world with synchronized players, block edits, and a reproducible local test flow.

Build a browser-based multiplayer voxel sandbox for [PLAYER COUNT] local or remote players. Implement a small generated world, first-person movement, block placement/removal, player names, synchronized transforms, and authoritative conflict handling for block edits. Use [NETWORK STACK] and keep the server setup minimal. Provide one command for the server and one for the client. Open two clients, verify join/leave behavior and state synchronization, and document known limits around persistence, latency, and cheating.

multiplayervoxel
AgenticExpert

Build And Self-Test An Apocalyptic Browser FPS

Uses a maximum-effort build-test-screenshot loop to deliver a playable FPS rather than a static visual demo.

Create a playable browser FPS set in an apocalyptic version of [CITY]. Include movement, aiming, shooting, enemies, health, ammunition, one objective, a win/lose state, recognizable environmental landmarks, and an accessible restart. Keep the level compact and use generated or procedural assets. Work autonomously at maximum useful effort. Launch the game, play through the objective, capture screenshots, inspect visual and gameplay failures, and iterate until the complete loop works. Report controls, performance assumptions, and anything you could not verify.

fpsself-testing
Frontend & Motion@chetaslua · X
Single promptAdvanced

Create Interactive Motion Graphics Entirely In Code

Creates a code-driven motion piece whose controls and composition remain useful even when animation is paused.

Build an interactive motion-graphics experience for [PRODUCT OR MESSAGE] entirely in code using [WEB STACK]. Create a polished 15- to 30-second sequence with a clear visual hierarchy, synchronized typography and effects, timeline controls, play/pause/restart, and at least two parameters the viewer can adjust. The paused composition must still read as a deliberate design. Use generated shapes and effects rather than stock footage. Verify timing, responsiveness, and reduced-motion behavior in the browser.

motion designinteractive
Frontend & Motion@nicky_sap · X
AgenticAdvanced

Research A Person And Build An Animated Personal Site

Combines source-backed research, information architecture, animation, and browser review in one end-to-end site workflow.

Research [PERSON OR CREATOR] using publicly available sources, then build a polished animated personal website that reflects their actual work. First produce a source table for every biographical claim, project, date, and quote. Then define the page narrative, visual system, and interaction plan before coding. Build the site in [STACK] with responsive sections for introduction, selected work, timeline, and contact. Launch it in a browser, review desktop and mobile screenshots, and iterate on hierarchy, animation restraint, accessibility, and factual accuracy. Do not invent missing facts.

researchpersonal site
Frontend & Motion@TokenGremlin · X
AgenticExpert

Refine A Black Hole Visualization With Screenshot Feedback

Uses repeated screenshots as observations for correcting rendering, camera, and composition problems in a black-hole simulation.

Build an interactive real-time black-hole visualization in [WEBGL STACK], then refine it through a screenshot feedback loop. Include a warped star field, accretion disk, gravitational lensing approximation, camera controls, exposure controls, and a short explanation of which effects are physically inspired versus artistic. After each working version, capture screenshots from the same three camera positions. Compare them against [REFERENCE OR TARGET], identify the largest visible mismatch, change one rendering subsystem at a time, and repeat until the remaining differences are documented.

webglvisual QA
Coding & Integrations@scottstts · X
AgenticExpert

Build A Virtual Laptop With A Functional Browser OS

Combines a 3D laptop shell with a usable browser desktop, working apps, and persistent interface state.

Build a virtual laptop experience that combines a Three.js hardware model with a functional desktop operating system rendered on its screen. The desktop must include movable windows, a dock, file explorer, notes, settings, a simple browser, and one media app. Implement window focus, minimize/maximize, keyboard input, and persistent local state. The laptop camera should orbit without breaking interaction with the screen. Test every app and window state, refresh persistence, mobile fallback, and the handoff between 3D camera controls and desktop pointer input.

threejsbrowser os
Coding & Integrations@rohanpaul_ai · X
Long contextExpert

Build A GPU Compiler From A DSL To PTX

Structures an end-to-end compiler task across parsing, intermediate representation, optimization, PTX emission, and verification.

Design and implement a small GPU compiler that translates [DSL] into NVIDIA PTX for [SUPPORTED OPERATIONS]. Work in phases: define the grammar and semantics; build the parser and typed IR; add validation and explicit error messages; implement at least two optimization passes; emit readable PTX; and add a Tensor Core path only where the operation and target support it. Create correctness tests that compare generated kernels with a trusted CPU reference, plus negative tests for invalid programs. Document architecture, supported syntax, unsupported cases, and how to inspect emitted PTX.

compilergpu
Research & Evaluation@tphuang · X
Long contextAdvanced

Research Multiple Topics From Chinese-Language Sources

Runs a long-form multilingual research workflow while keeping source quality, claim coverage, and latency visible.

Research these topics using primarily Chinese-language sources: [TOPIC LIST]. Produce one integrated analytical brief for [AUDIENCE]. Before writing, build a source ledger with URL, publisher, date, source type, relevant claim, and any reliability concern. Distinguish primary documents, reporting, commentary, and reposts. For each major conclusion, cite at least one direct source and note disagreement or missing evidence. Return the brief, the source ledger, unresolved questions, and a short process report covering search scope, time spent, and where long-running research produced diminishing returns.

deep researchChinese sources
Coding & Integrations@MinLiBuilds · X
AgenticAdvanced

Build A Synchronized Frontend Result Comparator

Adds a reusable comparison surface for selecting two completed frontend tasks and inspecting them side by side.

Add a dual-task comparison mode to [EXISTING FRONTEND EVALUATION APP]. Users must be able to select any two completed tasks, open their rendered outputs side by side, synchronize viewport size and scrolling, toggle interaction sync where safe, and inspect task metadata, screenshots, and evaluation notes without losing either view. Reuse the existing task data and components. Define a small comparison-state model, handle missing or failed outputs, add shareable URL state, and verify keyboard navigation, responsive fallback, and cross-frame security constraints.

evaluation tooliframe sync
Games & 3D@aimlapi · X
Single promptAdvanced

Build A Chameleon Hide-And-Seek Game In One Shot

Builds a self-contained color-matching game with procedural areas, sound, scoring, and a complete multi-round loop.

Build a complete chameleon hide-and-seek browser game as a single self-contained HTML file. The player controls a chameleon that blends into nearby colors to avoid [PREDATOR OR DETECTOR]. Include procedural hiding areas, readable color-matching feedback, movement, detection rules, sound effects generated in code, three rounds of increasing difficulty, scoring, game over, and restart. Use no external assets or build step. Play every round, verify that detection and scoring are fair, and fix any state that prevents a full restart.

single-file gameprocedural
AgenticExpert

Build A Cohesive 3D Wuxia Browser RPG

Combines traversal, melee combat, quests, inventory, weather, interiors, and exploration into one bounded browser RPG slice.

Build a cohesive 3D wuxia browser RPG vertical slice in [WEB STACK]. Include one village and one wilderness area, responsive traversal, melee combat with two techniques, one quest chain, inventory, weather, an enterable interior, NPC dialogue, and a final encounter. Use a consistent Chinese-fantasy visual language across architecture, clothing, effects, and UI. Implement systems in playable order, then complete the full quest yourself. Capture screenshots at each major location, fix blockers and obvious visual inconsistencies, and document controls plus unfinished expansion hooks.

wuxiarpg
Coding & Integrations@teortaxesTex · X
AgenticExpert

Build A Browser Game Console Around mGBA WASM

Integrates a real emulator core into a polished browser console while keeping ROM handling, controls, licensing, and state explicit.

Build a browser-based handheld game-console experience around the licensed mGBA WASM core. Create a polished console shell, ROM file picker, keyboard and gamepad mapping, pause/reset, save-state management, audio controls, fullscreen mode, and clear empty/error states. Do not bundle copyrighted ROMs. Preserve third-party notices and document the emulator core license. Test with a legally obtained homebrew ROM, verify input, audio, save/load, refresh behavior, and mobile fallback, then recursively improve the largest usability issue you observe.

wasmemulator
Frontend & Motion@1littlecoder · X
Single promptAdvanced

Create A Three.js Product Page From Two Reference Images

Uses two visual references with explicit roles to create a responsive 3D product presentation and conversion-ready page.

Create a responsive product page for [PRODUCT] using these two references: image 1 controls product shape and materials; image 2 controls art direction, typography, and layout. Build an efficient Three.js hero with a rotatable product, controlled lighting, detail hotspots, and a graceful static fallback. Add concise benefit copy, specifications, one primary CTA, and a mobile layout that preserves the product focus. Compare browser screenshots against both references, list deliberate deviations, and fix loading, interaction, contrast, and performance issues.

threejsproduct page
Frontend & Motion@filicroval · X
Single promptAdvanced

Invent A Luxury Product And Build Its Launch Page

Combines product invention, mechanism explanation, exploded-view storytelling, and a polished launch page in one artifact.

Invent a premium [EVERYDAY PRODUCT] for [TARGET CUSTOMER], then build its launch page. Define the product name, problem, mechanism, materials, key dimensions, and three credible benefits. Present a hero render, interactive exploded view, short use demonstration, specifications, care notes, and a clear CTA. Keep the design luxurious but technically legible. Label speculative engineering claims, check that every visual component matches the written mechanism, and verify the page on desktop and mobile.

product designexploded view
Frontend & Motion@1littlecoder · X
Single promptIntermediate

Generate A Seamless Recursive Story GIF

Builds a tightly timed ten-second narrative animation whose final frame returns naturally to the opening.

Create a seamless 10-second looping GIF about [CHARACTER] discovering [OBJECT OR EVENT]. Plan the animation as five two-second beats. The story must progress clearly, then use a visual recursion so the final frame matches the opening composition without a visible jump. Keep the character design, palette, camera, and lighting consistent. Render a preview, inspect the loop boundary frame by frame, and revise timing or composition until the transition is seamless. Deliver the GIF plus the editable source.

animationgif
Frontend & Motion@HarshithLucky3 · X
Single promptIntermediate

Generate A Detailed Vehicle Side-View SVG

Produces an editable side-view vehicle illustration with constrained proportions, layers, lighting, and clean vector structure.

Create an editable SVG side-view illustration of [VEHICLE] in [COLOR AND TRIM]. Match the recognizable wheelbase, roofline, body surfacing, wheels, windows, lights, and aero details while keeping the artwork original. Organize the SVG into named groups for body, glass, wheels, shadows, and highlights. Use gradients sparingly and avoid embedded raster images. Render the SVG at 320px and 1600px widths, fix clipping or unreadable detail, and report any proportions you approximated.

svgvehicle
Research & Evaluation@Satvik_Pen · X
Long contextAdvanced

Create A Post-Training Marketing PDF

Turns technical model-training material into a polished, source-aware marketing PDF for a defined audience.

Create a polished marketing PDF for [POST-TRAINING PRODUCT] aimed at [AUDIENCE]. Read the supplied research, evaluation results, product notes, and brand guidance. Build a source ledger before drafting. The PDF should explain the problem, method, evidence, workflow, limitations, and next step without inventing benchmark claims. Include one architecture diagram, one evidence table, and concise captions. Export the PDF, inspect every page as an image, fix overflow, weak hierarchy, tiny text, and unsupported claims, then deliver the final file plus a claim-to-source appendix.

pdfmarketing
Coding & Integrations@AlicanKiraz0 · X
Single promptExpert

Build A Real-Time Black Hole Raytracer In WebGL2

Builds a native WebGL2 geodesic-style raytracer with controls, performance instrumentation, and documented approximations.

Build a real-time black-hole raytracer as one HTML file using native WebGL2 and GLSL, with no rendering framework. Implement ray marching or geodesic-inspired bending, an accretion disk, Doppler-like color variation, gravitational lensing of a star field, orbit camera, quality controls, and an FPS display. Keep the shader readable and explain the math and approximations beside the demo. Test at multiple quality levels on [TARGET DEVICE], capture reference frames, and fix shader compilation, NaNs, severe banding, and controls before delivery.

webgl2raytracing
Coding & Integrations@twid · X
AgenticExpert

Clone A Desktop OS In The Browser With Working Apps

Creates a coherent browser desktop with shared window behavior and functional music, browser, mail, and settings apps.

Build a desktop-OS-style web app inspired by [DESIGN ERA], without copying protected branding or bundled assets. Implement a menu bar, dock, desktop icons, notifications, themes, and reusable window management. Include working music, browser, mail, file, and settings apps with meaningful local interactions and shared persistent state. Support focus, drag, resize, minimize, maximize, and keyboard navigation. Create an app-by-app test checklist, verify cross-app state and refresh persistence, and fix inconsistent window behavior before polishing visual effects.

browser osapps
Research & Evaluation@emollick · X
Long contextExpert

Run A Statistical Audit With Independent Review

Pairs a model-generated statistical audit with a separate review step before any finding is treated as decision-ready.

Audit the statistical analysis in [PAPER, REPORT, OR NOTEBOOK]. Reproduce the reported calculations where data and code permit. Check estimand, sampling, missing data, leakage, model assumptions, uncertainty, multiple comparisons, robustness, and whether conclusions exceed the evidence. For every finding, cite the exact table, code cell, page, or variable and assign a confidence level. Do not finalize recommendations. Produce a review packet for an independent statistician or second model: claims, evidence, reproduction steps, disputed points, and questions that must be resolved before relying on the audit.

statisticsaudit
Research & Evaluation@doodlestein · X
Long contextExpert

Audit A Complex Plan And Challenge Its Remedies

Uses a second-pass adversarial review to find understated risks, incorrect remedies, and conclusions that should be rejected.

Act as an independent reviewer of this heavily refined plan: [PLAN]. Assume the authors are competent, but do not assume prior review made it correct. Map every objective, assumption, dependency, risk, mitigation, owner, and success metric. Then challenge the plan in three passes: understated severity, remedies that do not address root causes, and conclusions unsupported by evidence. Cite the exact section for every criticism. Return a decision table with accept, revise, or reject; explain the minimum correction; identify disagreements with prior reviewers; and clearly separate verified defects from judgment calls.

plan auditsecond opinion

How To Write Better Kimi K3 Prompts

Define a finished artifact and a testable completion state, not just a visual theme.
For agentic builds, require a build-run-observe-fix loop with screenshots or tests.
Separate core functionality from second-pass polish so the first version can converge.
Name tools, sources, constraints, and uncertainty boundaries instead of letting the model infer them.

Kimi K3 Prompt FAQ

What are these Kimi K3 prompts based on?

Each template is adapted from a public Kimi K3 case in the linked GitHub collection. Every card names the source author and links to the original post. When the original full prompt was not public, this page presents a reusable template rather than claiming to reproduce the author’s exact wording.

What is Kimi K3 good at in these examples?

The source set shows strong interest in browser games, Three.js and Blender workflows, motion-heavy frontend work, long-running coding tasks, integrations, and research. Individual posts are demonstrations or evaluations, not universal benchmark results.

How should I prompt Kimi K3 for a long agent task?

Give it a bounded definition of done, the tools it may use, an explicit verification loop, and a reporting format for assumptions and unverified work. Add time or iteration limits for unattended runs.

Are these the authors’ original prompts?

Only where a public source exposes the prompt. Most cards are carefully written reusable templates adapted from the demonstrated objective, with the original author and post linked for context.

How do I run these prompts through EvoLink?

Copy a template, replace the bracketed variables, and send it to the `kimi-k3` route through EvoLink’s OpenAI-compatible Chat Completions API. Current pricing and access details remain on the Kimi K3 model page.