Kimi K3 现已上线查看 Kimi K3
Claude Fable 5 prompt gallery hero background

Claude Fable 5 Prompts

取自真实 Claude Fable 5 使用案例的可复用 prompt 模式,覆盖编程、agent 自动化、交互构建、前端设计和长上下文研究。复制即用。

浏览基于真实 Claude Fable 5 使用案例整理的 prompt 模板,选好后可跳转到模型页在 EvoLink 上运行。

精选 Claude Fable 5 Prompt 示例

一批持续扩充的可复用 prompt。想看更多真实用例,可以去 GitHub 指南

Coding & Refactoringcoldopn on X
AgenticAdvanced

从截图复刻应用界面

把一张现有产品的截图变成一个可运行、有实际功能的界面克隆。

Here is a screenshot of [PRODUCT NAME]'s interface: [ATTACH SCREENSHOT]. Recreate this UI as a working [FRAMEWORK / SINGLE HTML FILE] app with REAL functionality, not a static mockup. - Match the layout, spacing, and visual hierarchy in the screenshot as closely as you can. - Implement the core interactions I would actually use: [LIST KEY FEATURES, e.g. navigation, search, create/edit items]. - Wire up local state so buttons, forms, and lists genuinely respond. Work autonomously until it runs, then tell me what you approximated and what still needs real backend wiring.

reverse-engineeringui-clone
Coding & Refactoringmeta_alchemist on X
Long contextAdvanced

仓库审计与改进计划

以资深工程师视角审计整个仓库,返回带文件证据、按优先级排序的改进计划。

You are a principal-level software engineer and technical auditor. Deeply analyze this repository and deliver an honest audit plus a prioritized, actionable improvement plan. Ground every claim in actual files: cite file paths and line numbers. If you cannot verify something, say so. Work through these phases in order, do not skip ahead: 1. Map the architecture, key modules, and how data flows through [PROJECT NAME]. 2. Find correctness bugs, security issues, and [SPECIFIC CONCERN, e.g. performance hot paths]. 3. Flag dead code, duplication, and risky patterns worth refactoring. 4. Produce a ranked improvement plan (impact vs effort) with concrete next steps.

auditrefactoring
Coding & RefactoringPilgrim6938 on X
Long contextExpert

在大型代码库中定位并修复缺陷

在大型仓库中追踪疑难缺陷,定位真正的根因,并给出可直接进入评审的修复方案。

This repository has roughly [LINE COUNT] lines. There is a bug: [DESCRIBE THE SYMPTOM, e.g. cache never invalidates after an update]. Do a full investigation before proposing anything: - Trace the relevant call chains end to end, including hidden or indirect calls. - Identify the true root cause, not just the surface symptom, and cite the files and lines involved. - Propose a complete fix that a reviewer could merge, covering edge cases and any affected callers. Explain your reasoning and list anything you were unsure about so I can verify before merging.

debugginglarge-codebase
Agents & Automationayushagarwal on X
AgenticExpert

排查线上故障并提交修复 PR

跨日志与多个服务排查线上故障,定位根因,并自动提交修复 PR。

Something is broken in [ENVIRONMENT, e.g. my self-hosted k8s cluster] and I cannot describe it precisely. Investigate and fix it. You have access to [AVAILABLE TOOLS, e.g. kubectl, database query access, container registry]. Work autonomously: - Read pod/service logs and error logs from [DATA SOURCES] to find what actually changed. - Compare deployed versions, image digests, config, and migrations to spot mismatches. - Form a hypothesis, verify it against evidence, then open a PR with the fix and a clear write-up. Do not ask me to hand-hold each step; report your findings and the PR when done.

incident-responsedevops
Agents & Automationoutsource_ on X
AgenticExpert

自动化构建-测试-运行闭环

搭建一个无人值守的循环,让智能体持续构建、测试并冒烟验证,直到项目完成。

Build [PROJECT NAME] to completion by running an autonomous loop, hands-free, until every item in the spec is met. The spec / definition of done is: [SPEC OR CHECKLIST]. On each iteration: - Implement the next unfinished piece. - Build it, run the tests, and smoke-test the actual behavior (launch it, exercise it, [e.g. play the level]). - If something fails, diagnose and fix it, then re-verify before moving on. Keep a running log of what is done, what changed, and what remains. Continue on autopilot until the checklist is fully satisfied.

autonomouslong-horizon
Agents & Automationoikon48 on X
AgenticAdvanced

目标驱动、自我验证的工作流

把任务设为一个目标让模型持续推进直到完成,再逐项验证并报告与计划的差异。

Set a goal to fully implement this spec, then keep working until it is genuinely complete: [SPEC]. Context you should know: [BACKGROUND AND CONSTRAINTS, e.g. this is experimental and may be removed in a month, so keep it easy to delete]. Then run a workflow that verifies each part of the plan independently: - For every requirement, check the real behavior against what was asked. - Note anything that differs from the spec or that you had to interpret. Prepare a short report of what was implemented, how you verified it, and where reality diverged from the plan.

goalworkflow
Games & Interactiveai_for_success on X
Single promptIntermediate

手势控制的跑酷游戏

构建一个通过摄像头手指/手势追踪操控的浏览器跑酷游戏,单文件即可运行。

Build a browser-based endless-runner game controlled by [INPUT, e.g. finger/hand tracking via the webcam]. Requirements: - Use [LIBRARY OR APPROACH, e.g. a hand-tracking library] to map [GESTURE, e.g. raising a finger] to the jump action. - Add obstacles, increasing speed, a score counter, and a game-over/restart flow. - Keep it self-contained and runnable as a single [HTML file / small project] with no build step. Handle camera permission gracefully and keep it at a smooth frame rate.

gamewebcam
Games & Interactivekieradev on X
Single promptAdvanced

一次生成的卡丁车竞速游戏

用一句简短提示生成含赛道、角色、模式、音乐和 UI 的完整街机卡丁车游戏。

Build a complete arcade kart-racing game in the style of [REFERENCE, e.g. Mario Kart 64]. Include, generated by you, no external assets: - At least [N] tracks with distinct themes and decoration. - Driveable characters/karts with simple models and handling. - [GAME MODES, e.g. grand prix, time trial, versus], plus background music and sound. - A full UI: menus, HUD, lap and position tracking, and results screen. Keep controls responsive and ship it as a runnable project I can start immediately.

game3d
Games & Interactiveydamitcodes on X
AgenticAdvanced

体素沙盒游戏克隆

用一条提示构建含地形生成、挖掘、昼夜循环和生物的可玩体素沙盒游戏。

Build a playable voxel sandbox game in the style of [REFERENCE, e.g. Minecraft]. Include: - Procedurally generated terrain with [FEATURES, e.g. biomes, caves, ores]. - Block placing and mining with a simple inventory. - A day/night cycle and basic [MOBS/ENTITIES] with simple AI. - First-person movement, collision, and mouse-look controls. Target [TECH, e.g. a single HTML file with WebGL / a Rust project] and keep a steady frame rate. Build it end to end, testing as you go.

gamevoxel
Frontend & Designintheworldofai on X
Single promptAdvanced

单个 HTML 文件里的桌面操作系统

用一个 HTML 文件生成带窗口、程序坞和可用应用的完整桌面操作系统体验。

Build a full [STYLE, e.g. macOS-style] desktop operating system as a single self-contained HTML file with embedded CSS and JS. Include: - A menu bar and dock, plus draggable and resizable app windows. - Working built-in apps: [APPS, e.g. file explorer, terminal, settings, browser mockup]. - Desktop icons, notifications, themes, and a boot screen. - [BONUS, e.g. a small working sound player using real synthesis, or a mini game inside the desktop]. Make it feel real and responsive, generated in one pass with no external files.

uisingle-file
Frontend & Designecomchasedimond on X
Single promptIntermediate

高端电商邮件视觉稿

一次生成一封逼真、竖向长图的高端电商品牌营销邮件设计稿。

Create a full-length, tall vertical ecommerce email mockup for [BRAND], a [CATEGORY] brand for [AUDIENCE]. Design a highly realistic premium DTC email that feels sleek, polished, and editorial, like a real designer-made email screenshot, not a poster, landing page, or generic template. Brand vibe: [ADJECTIVES, e.g. modern, confident, clean, aspirational]. Feature these products prominently: [PRODUCT LIST]. Include a hero, product sections with real-feeling copy, and a clear call to action. Use realistic typography, spacing, and imagery placement.

designemail
Frontend & Designscottstts on X
Single promptAdvanced

按平面图生成可漫游的 3D 场景

用 Three.js 依据给定平面图构建可第一人称漫游、忠实还原布局的 3D 房间。

In Three.js, create a first-person, navigable 3D scene of [PLACE, e.g. the iconic apartment from a TV show] based on this floor plan: [ATTACH FLOOR PLAN]. - Stay true to the original look and feel: layout, furniture placement, colors, and materials. - Use [LIGHTING STYLE, e.g. warm interior lighting] and sensible shadows. - Let me walk around with WASD and mouse-look, with collision so I cannot pass through walls. Deliver it as a single HTML file. Verify your result visually against the floor plan and iterate until the geometry matches.

3dthreejs
Docs & Researchmanaltdan on X
Long contextIntermediate

对长文档做修订标记式审阅

通读一份长文档,返回修订标记式编辑、对存疑论点的质询以及可补充的资料来源。

Do a full, careful read-through of this [DOCUMENT TYPE, e.g. policy memo]: [PASTE OR ATTACH DOCUMENT]. Return your review in three layers: - Fluency edits: tighten wording and flow, shown as clearly marked track changes. - Comments: flag shaky or unsupported claims and ask the questions a sharp reviewer would. - Sources: suggest specific references or data that would strengthen or verify the weak points. Preserve my voice and argument; do not rewrite it into your own style. Keep the changes suggestion-style so I can accept or reject each one.

editingreview
Docs & Researchnaoyafujiwara on X
Long contextIntermediate

结构化深度研究简报

针对一个宏大开放问题,产出既有深度又结构清晰的分析简报。

Research and answer this question in depth: [QUESTION]. Frame it around a specific premise: [PREMISE OR ANGLE, e.g. treat the timeline as starting from a particular year rather than the conventional one]. Structure the answer as: - What actually happened: the concrete events, decisions, and turning points. - Why it happened: the underlying causes and how they compounded. - What it means going forward: implications for [TIME HORIZON / DOMAIN]. Go deep on substance but keep the surface clean and readable. Cite specific figures, dates, and facts, and flag where the evidence is contested.

researchanalysis
Docs & Researchgabepereyra on X
Long contextAdvanced

起草 S-1 上市登记文件

为指定公司起草结构规范、篇幅完整的 S-1 上市登记文件初稿。

Draft an S-1 registration statement for [COMPANY], a [DESCRIPTION] company. Follow the standard S-1 structure and produce genuinely long-form content, not an outline: - Prospectus summary, risk factors, use of proceeds, and MD&A. - Business description, competition, and market opportunity. - Selected financial data and management/ownership sections. Match the tone, formatting, and structure of a real filing. Where you would need actual numbers, use clearly labeled placeholders such as [FY2025 REVENUE] rather than inventing figures.

legalfinance

如何写出更好的 Claude Fable 5 Prompt

先写目标,再写约束——保留什么、不要动什么、允许用哪些依赖。
复杂任务先让它出方案再写代码,这样能低成本地拦下错误思路。
要求验证:测试、类型检查或实际运行——"看起来完成"不等于完成。
让它主动标注假设和模糊点,而不是默默猜测。

Claude Fable 5 Prompt 常见问题

Claude Fable 5 最擅长什么?

Fable 5 是 Anthropic 目前最强的模型,档位在 Opus 之上,拥有业界领先的编程、推理和 agent 能力,以及 100 万 token 上下文窗口。它在多文件编程、长链路 agent 运行、重构和超长文档综合上表现突出。

这和图像 / 视频提示词页有什么不同?

Fable 5 是文本、编程和 agent 模型,所以这些 prompt 是可以粘贴到对话、编程 agent 或 API 里的可复用指令,而不是媒体生成提示词。每张卡片展示完整 prompt 并说明它为什么有效。

怎么写好一个 Fable 5 prompt?

把目标和约束分开写,复杂任务先要方案再写代码,要求明确的验证,并让模型标注假设而不是猜。本页的框架部分对此有拆解。

这些 prompt 能在 Claude Code 或我自己的工具里用吗?

可以。这些是与具体模型无关的 prompt 模式。用 EvoLink,你可以把 Claude Code 或任意 OpenAI 兼容客户端指向我们的接口,用同样的 prompt 运行 Fable 5。

这些 prompt 怎么配合 EvoLink 使用?

点击卡片上的复制,把 [占位符] 换成你自己的内容,然后通过 EvoLink 在 Fable 5 上运行——可以走 API、Claude Code,或任意指向我们接口的 OpenAI 兼容客户端。定价和接入方式见模型页。