Mobiagent Task Spec

Use this skill whenever a task involves MobiAgent or steward mobile execution. It defines how to write high-precision mobile task descriptions, including goal + step-level click/swipe/input instructions, explicit completion criteria, and evidence-driven rewrites based on final screenshot + status/reasoning + VLM extraction evidence. Trigger for requests about mobile automation success rate, steward task wording, evidence-based correction, or improving phone-side accuracy.

43 stars 1 forksProductivity & Notes

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add IPADS-SAI/MobiClaw --skill mobiagent-task-spec
Manual install
git clone --depth 1 https://github.com/IPADS-SAI/MobiClaw.git /tmp/MobiClaw
mkdir -p ~/.claude/skills
cp -r /tmp/MobiClaw/mobiclaw/skills/mobiagent-task-spec ~/.claude/skills/mobiagent-task-spec

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# MobiAgent Task Spec

用于把“模糊手机任务”改写成 MobiAgent 可执行、可验证的描述,并指导 steward 正确消费单次执行结果。

## 1) 能力边界(必须先遵守)

- MobiAgent 本质是基础动作组合执行器:`点击`、`滑动`、`输入`(以及导航过程中的等价基础操作)。
- 不能把它当作“高级意图 API”来用。避免要求“总结”“截图”“导出报告”“直接读取某字段”这类一步到位动作。
- 任务描述尽可能落到页面元素与动作序列上,否则会漂移或死循环。

## 2) 任务描述原则

- 目标单一:每次下发手机执行的任务只涉及1个APP,只针对1个完整任务。
- 步骤可执行:每一步都写“看到什么 -> 做什么”。
- 完成可判定:必须有可观察的成功标准,不用“应该/大概/尽量”。
- 约束可控:写清副作用约束(例如“不要发送消息,只停留在确认页”)。

## 3) 标准任务模板(自然语言版,给 steward 直接用)

`task_desc` 只写一段自然语言,不要使用分段标题,也不要分点罗列。  
写法要求:把核心目标、关键步骤、失败分支、成功判定、禁止项合并为连续描述。  
重试时也一样:`task_desc` 里只保留“修正后的任务描述正文”,不要包含“失败原因/修正点/改写说明”。

可直接套用下面句式并替换占位符:

```text
请在<App>中完成以下任务:先确保当前位于<App>首页,如果不在则回到首页;然后定位<锚点A>并点击进入<页面B>,如果没看到<锚点A>就先滑动一次再重试,连续两次仍找不到就回退到上一级页面后重新从首页开始;进入<页面B>后定位<锚点C>,必要时在搜索框输入<关键词>并进入目标页面,最终停留在包含<锚点D>的页面。任务完成标准是同时满足<页面级锚点>和<目标对象锚点>可见,并且屏幕中能看到<证据关键词>;执行过程中不要进行<禁止动作1>和<禁止动作2>等与目标无关操作。
```

任务描述 `task_desc` 示例:
- 简易版:打开爱奇艺浏览最近的电影预告。
- 详细版:使用爱奇艺应用查看新片预告,具体操作步骤为:1. 进入爱奇艺首页后,连续向上滑动屏幕浏览页面内容;2. 在页面下方找到“新片即将上线”的相关板块;3. 点击该板块中的“更多新片预告”按钮,即可进入详情页面查看即将上线的新片信息。

## 4) 好坏对比(必须避免坏例子)

坏例子(不可执行):
- “帮我把微信最近消息整理一下并截图。”
- “去支付宝看下账单然后告诉我重点。”

好例子(可执行):
- “请在微信中找到‘项目A群’聊天页并停留:先进入微信首页,点击顶部搜索输入‘项目A群’后进入聊天页;完成标准是页面顶部显示‘项目A群’且聊天区可见最近消息文本,不要发送任何消息。”
- “请在支付宝中进入本月账单列表页并停留:先打开支付宝并进入‘账单’,再筛选‘本月’,连续失败则返回账单首页重试;完成标准是页面同时出现‘账单’和‘本月’文本且可见金额(如¥),不要进行支付或转账操作。”

## 5) 重试改写流程(基于已有证据)

当执行失败后,必须读取并利用以下证据再改写任务:
- 最后一张截图(ImageBlock / `final_image_url`)
- `last_reasoning`
- `status_hint`
- `vlm_summary_screen_state` / `vlm_summary_last_steps`
Read the full skill on GitHub

Details

Skill name
mobiagent-task-spec
Path
mobiclaw/skills/mobiagent-task-spec/SKILL.md
Repo updated
27 March 2026

FAQ

What does the Mobiagent Task Spec skill do?

Use this skill whenever a task involves MobiAgent or steward mobile execution. It defines how to write high-precision mobile task descriptions, including goal + step-level click/swipe/input instructions, explicit completion criteria, and evidence-driven rewrites based on final screenshot + status/reasoning + VLM extraction evidence. Trigger for requests about mobile automation success rate, steward task wording, evidence-based correction, or improving phone-side accuracy.

How do I install the Mobiagent Task Spec skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add IPADS-SAI/MobiClaw --skill mobiagent-task-spec. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Mobiagent Task Spec popular?

It lives in IPADS-SAI/MobiClaw, which has 43 GitHub stars and 1 forks.

Related skills

Claude API

43

IPADS-SAI/MobiClaw

Build apps with the Claude API or Anthropic SDK. TRIGGER when: code imports `anthropic`/`@anthropic-ai/sdk`/`claude_agent_sdk`, or user asks to use Claude API, Anthropic SDKs, or Agent SDK. DO NOT TRIGGER when: code imports `openai`/other AI SDK, general programming, or ML/data-science tasks.

AI, LLMs & Agents

Setup

43

IPADS-SAI/MobiClaw

Run initial MobiClaw setup. Use when user wants to install dependencies, configure environment variables, or start the gateway server. Triggers on "setup", "install", "configure", "start gateway", or first-time setup requests.

Other

Windows UI Automation

48

martinholovsky/claude-skills-generator

Expert in Windows UI Automation (UIA) and Win32 APIs for desktop automation. Specializes in accessible, secure automation of Windows applications including element discovery, input simulation, and process interaction. HIGH-RISK skill requiring strict security controls for system access.

Productivity & Notes

Celery Expert

48

martinholovsky/claude-skills-generator

Expert Celery distributed task queue engineer specializing in async task processing, workflow orchestration, broker configuration (Redis/RabbitMQ), Celery Beat scheduling, and production monitoring. Deep expertise in task patterns (chains, groups, chords), retries, rate limiting, Flower monitoring, and security best practices. Use when designing distributed task systems, implementing background job processing, building workflow orchestration, or optimizing task queue performance.

Productivity & Notes

Act.quickstart

57

javimosch/supercli

Agent workflow for running, listing, validating, and testing GitHub Actions workflows locally via act CLI.

Productivity & NotesIn 1 list

Browser Automation

48

martinholovsky/claude-skills-generator

Expert in browser automation using Chrome DevTools Protocol (CDP) and WebDriver. Specializes in secure web automation, testing, and scraping with proper credential handling, domain restrictions, and audit logging. HIGH-RISK skill due to web access and data handling.

Productivity & Notes

Browse the most popular Claude skills, all productivity & notes skills, or write your own with the skills generator.