本文来源:https://mp.weixin.qq.com/s/36Dk2TEcijpcpOrJpHjyqg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

作者背景介绍:
-
个人职业:美妆行业的数据分析技术岗
-
日常高频工作内容:经营数据采集、数据清洗、数据分析、数据报告提报
-
本次使用功能:TRAE Work + Skill 自动化 + 飞书云文档
前言

作为美妆行业的数据分析岗,我的日常离不开多数据源的周期性采集、清洗、整合与提报,长期依赖人工复制粘贴和 RPA,链路割裂、易出错、纯耗时。
本文我将为大家分享如何用 「TRAE Work + Skill 自动化 + 飞书云文档」,把这条模糊、易错、依赖人工经验的办公流程,逐步改造成一条可执行、可验证、可调度、可协作的数据生产线:从封装采集 Skill、配置定时任务,到自动归档飞书、数据自动校验,最终实现每周一数据自动出现在飞书、开箱即评审的无人值守闭环。
文中同时复盘了 HID 弹窗、单位丢失等真实难题的解法,以及我的使用技巧与改进期待。

实践背景
在做这套自动化之前,我的经营数据工作长期卡在「链路长、环节割裂、稳定性差」这几个痛点上。
整个流程从采集到提报要跨多个环节,每一环都要人工介入,既耗时又容易出错。具体来说:
-
【需求高频且周期性】在运营数据需求支持工作中,需要定期对多个数据源进行周期性趋势定位与分析,才能充分判断运营策略是否需要调整和把控节奏。
-
【全链路割裂】数据采集、清洗、整合、分析、报告提报,以往都是各自割裂、彼此脱节的。
-
【采集不稳定】刚开始数据采集靠手工一条条复制粘贴;后来改用 RPA 软件模拟采集,人虽然轻松一些,但经常定位不准、突然卡断,往往要重跑好几次才能拿到结果。
-
【清洗标准不统一】数据清洗与整合第一阶段标准不一致,只能靠人眼手工处理;第二阶段虽用脚本快速清洗合并有所改善,但采集与清洗、整合之间仍是割裂分离的状态。
-
【提报难追溯】数据提报只能通过飞书消息对话传文件,无法很好地形成追溯与沉淀。


核心场景
场景一:Skill 驱动的数据采集(替代人工操作网页)
我把自己日常在数据平台上的完整操作流程,封装成了一个 TRAE Skill(data-automation)。
整个 Skill 包含一条完整的 CLI 自动化流水线:
环境检查 → 依赖安装 → 经营概览采集 → 条件筛选 → 经营明细数据采集→ 数据清洗 → 数据合并 → 数据校验 → 输出最终 Excel
在 TRAE Work 中只需要一句话就能启动:
请执行 data-automation 完整流程
TRAE Work 会自动读取 Skill 说明、安装依赖、打开 Chrome、完成登录授权、逐一采集数据、清洗合并、校验通过后输出 Excel。

场景二:TRAE Work 定时任务(无需人工关注)
数据采集不需要我每天手动触发。我通过 TRAE Work 的定时任务功能,配置了每周一早上 7 点自动执行:
-
执行 data-automation Skill 完整流程
-
将合并后的结果文件上传飞书云文档
配置完成后,每周一早上到工位打开飞书,最新的数据报告已经生成好了。

场景三:飞书云文档自动归档(团队协作更敏捷)
过去采集完数据后,还要手动上传到飞书、更新文档链接、通知团队成员。现在这一步也自动化了。
Skill 执行完成后,TRAE Work 会自动调用 飞书 CLI,将最终 Excel 作为飞书电子表格上传到飞书云文档。团队成员打开飞书就能直接查看最新数据,周一上班直接打开数据结果链接即可开会评审。
整个流程从“人工操作网页 → 手动整理 Excel → 手动上传飞书”,变成了:
“TRAE Work 定时触发 → Skill 自动采集 → 自动清洗合并校验 → 自动上传飞书电子表格 → 团队直接在飞书查看”

*图中涉及的数据敏感信息已屏蔽 / 修改
场景四:数据校验与质量保障(不让错误流入下一环节)
这是和之前的工作流最大的区别。以前人工操作时,漏选一个筛选条件可能几天后才发现,但是现在每一步都有校验:
-
每个核心采集点必须全部采集成功
-
每个核心采集点必须有 30 条数据
-
单位转换需兼顾中文可读性,数值换算与排序需同时支持
-
最终 Excel 必须包含品牌达人明细和品牌概览两个 Sheet
-
任一校验不通过,流程立即停止并保存 debug 截图

如何实践
我没有一开始就让 AI “直接写一个完整脚本”,而是把工作拆成多个可以肉眼验收的步骤。
第一步:先让 TRAE Work 理解真实流程
我把原有 RPA 流程图、Python 脚本、操作录屏和字段要求一起交给 TRAE Work,让它先整理流程,而不是急着写代码。
先不要完整运行。请根据 操作步骤 流程图、现有 Python 脚本和操作录屏,整理从登录、采集、明细采集,到清洗和合并的完整操作步骤,并列出每一步的输入、输出和校验条件。

*涉及数据敏感信息已屏蔽/已修改
第二步:分段执行,逐段验收
整个项目按阶段推进:登录验证 →品牌概览 → 全量概览 → 筛选验证 → 单达人 → 全量达人 → 清洗测试 → 合并测试 → 完整流程 → 飞书上传 → 定时任务配置。
这个过程最有价值的一点,是 TRAE Work 可以持续读取运行日志、截图和 Excel 结果,再修改脚本,而不是每次都从头解释项目。
第三步:从“能跑”升级到“能调度”
脚本稳定后,我让 TRAE Work 帮我做了三件事:
-
封装成 Skill:补齐 SKILL.md、配置文件、自动依赖安装、默认输出目录、结果校验,让其他环境也能调用;
-
配置定时任务:通过 TRAE Work Work 的 Schedule 功能,设定每周一自动执行;
-
对接飞书:通过飞书 CLI,将结果自动上传为飞书电子表格。

TRAE Work 如何解决最难的问题
难题一:Chrome 原生 HID 弹窗无法用网页选择器控制
登录后需要点击右上角头像、点击“更新”,再在 Chrome 原生 WebHID 浮窗中选择 USBKey。
最初只能使用屏幕坐标点击,但这对分辨率和系统缩放非常敏感,不适合打包成可迁移的 Skill。
最终方案:
-
枚举 USBKey 的 VID/PID;
-
为目标网站写入 Chrome WebHidAllowDevicesForUrls 策略;
-
默认关闭坐标点击,仅在策略无法生效时保留鼠标点击作为兜底。
难题二:“亿”单位丢失导致排名错误
1.26亿 被截成 1.26,3.21亿 被截成 3.21,亿元级数据反而排到了万元级后面。
TRAE Work 帮我从最终 Excel 一路反查到采集源头,发现销售额拆分正则只识别 w,没有识别中文的 万/亿。最终处理方式:采集阶段完整保留 万、亿、w,展示值保留原单位、排序时统一换算为数值、校验时主动查找疑似被截断的裸数字。

*涉及数据敏感信息已屏蔽 / 已修改

最终成果
最终成果是一个 Skill + 定时任务 + 飞书云文档 的完整自动化方案:
data-automation-skill/├─ SKILL.md # Skill 说明文件├─ run_full_pipeline.ps1 # 单一入口├─ config/│ ├─ config.json # 运行配置│ └─ config.example.json # 配置模板├─ scripts/│ ├─ data_brand_scraper.mjs # 概览采集│ ├─ data_blogger_scraper.mjs # 明细数据采集│ ├─ clean_blogger_excel.py # 数据清洗│ ├─ merge_feigua_data.py # 数据合并│ ├─ setup_chrome_webhid_policy.ps1 # WebHID 策略│ └─ validate_pipeline_outputs.py # 结果校验└─ output/ # 输出目录
三层自动化架构:
|
层级 |
工具 |
职责 |
|
执行层 |
Skill (data-automation) |
数据采集、清洗、合并、校验 |
|
调度层 |
TRAE Work 定时任务 |
每周一 07:00 自动触发执行 |
|
协作层 |
飞书 CLI + 云文档 |
结果自动上传为飞书电子表格 |

效率提升对比
时间对比
|
环节 |
人工操作 |
自动化后 |
|
概览采集 |
~60 分钟 |
AI执行我在睡觉 |
|
明细采集 |
~1.5 小时 |
AI执行我地铁 |
|
数据清洗与合并 |
~30 分钟 |
|
|
上传飞书云文档 |
~2 分钟 |
|
|
合计 |
~3 小时+人工值守 |
AI执行没参与+无人值守 |
质量提升
以前人工操作中,漏选筛选条件、分页遗漏、单位丢失等问题时有发生,往往开会发现错漏,临时对数据需要返工。
现在:AI 执行,我直接提交结果
工作方式的变化
以前:每周一早上手动打开数据平台 → 逐一搜索 → 切换页面 → 复制字段 → 翻页 → 清洗 → 合并 → 上传飞书 → 通知团队
现在:每周一早上打开飞书,数据已经在那里了。

5 个 TRAE Work 使用技巧
技巧 1:先让 AI 复述流程,再让它写代码
如果 AI 连真实操作路径都没有理解,写得越快,返工越多。
技巧 2:把“肉眼看到的问题”转成可执行校验
不要只说"推广按钮好像没点到",而要改成:"先判断“推广”按钮位置,再尝试操作,仍未检测到推广按钮时,停止采集并保存截图。"问题不会在下一次执行中悄悄复发。
技巧 3:单核心采集点验证通过后,再跑全量
全量运行接近 50 分钟。如果不先做单核心采集点验证,任何小错误都会浪费一整轮时间。
技巧 4:让 AI 检查最终文件,而不只是检查“脚本”
真正的完成标准不是终端显示成功,而是 Excel 文件存在、行数正确、字段没有错位、金额单位没有丢失、排序符合业务逻辑。
技巧 5:学会用 Skill + 定时任务 + 飞书实现“闭环自动化”
脚本稳定后,分三步升级你的工作流
-
封装成 Skill:让 AI 可直接调用
-
配置定时任务:让执行可调度
-
对接飞书云文档:让结果可协作
这一步让项目从"我的电脑上能跑",变成"每周一早上数据自动出现在团队的飞书里"。

写在最后
这次实践让我感受最深的是:AI 编程真正有价值的地方,不只是帮我多写了几百行代码,而是帮助我把一个模糊、易错、依赖人工经验的办公流程,逐步变成了可执行、可验证、可调度、可协作的系统。
从最开始“怎么点到某个按钮”,到“怎样让其他 AI 也能执行 Skill”,再到“怎样让数据每周一自动出现在飞书里”:项目的目标在不断升级。
最终交付的不只是一个 Excel,而是一套能稳定生产数据、自动归档到团队协作空间的无人值守生产线。

Source: https://mp.weixin.qq.com/s/36Dk2TEcijpcpOrJpHjyqg If you have any copyright issues, please contact: 13951431913#139.com or leave a comment on the right.
Author: Damond, a core TRAE user
Author Introduction:
Occupation: Data analyst in the beauty industry
Daily high-frequency work content: Business data collection, data cleaning, data analysis, and report submission
Feature used this time: TRAE Work + Skill Automation + Feishu Cloud Docs
Preface
As a data analyst in the beauty industry, my daily work revolves around periodic collection, cleaning, integration, and reporting across multiple data sources. I've long relied on manual copy-paste and RPA, which leads to fragmented workflows, errors, and pure time waste.
In this article, I'll share how to use 「TRAE Work + Skill Automation + Feishu Cloud Docs」 to gradually transform this vague, error-prone, experience-dependent office process into an executable, verifiable, schedulable, and collaborative data production line: from encapsulating collection Skills and configuring scheduled tasks, to auto-archiving in Feishu and automatic data validation, ultimately achieving an unattended weekly loop where data automatically appears in Feishu ready for review.
The article also reviews solutions to real challenges such as HID pop-ups and unit loss, along with my usage tips and improvement expectations.
Background
Before building this automation, my business data work was long stuck on pain points of "long chain, fragmented links, and poor stability."
The entire process from collection to reporting spans multiple stages, each requiring manual intervention, which is both time-consuming and error-prone. Specifically:
【High-frequency and periodic demand】 In operational data support work, periodic trend positioning and analysis across multiple data sources are needed to fully assess whether operational strategies require adjustment and to control the pace.
【Fully fragmented chain】 Data collection, cleaning, integration, analysis, and report submission were all previously isolated and disconnected from each other.
【Unstable collection】 Initially, data collection relied on manually copying and pasting one by one; later switched to RPA software simulation, which eased the burden but frequently suffered from inaccurate positioning and sudden interruptions, often requiring multiple re-runs to get results.
【Inconsistent cleaning standards】 The first phase of data cleaning and integration lacked uniform standards and had to be handled manually by eye; the second phase improved somewhat with scripts for fast cleaning and merging, but collection, cleaning, and integration remained in a disconnected state.
【Hard to trace submissions】 Data submission could only be done by sharing files through Feishu message conversations, making proper traceability and沉淀 difficult.
TRAE Work made me think of solving all these pain points in one go: through a truly unattended, auto-executing, self-validating data production line with results automatically archived in Feishu.
Core Scenarios
Scenario 1: Skill-driven data collection (replacing manual web operations)
I encapsulated my complete daily operation process on the data platform into a TRAE Skill (data-automation).
The Skill includes a full CLI automation pipeline:
In TRAE Work, just one sentence is enough to kick it off:
TRAE Work will automatically read the Skill description, install dependencies, open Chrome, complete login authorization, collect data one by one, clean and merge, validate, and output an Excel file.
Scenario 2: TRAE Work scheduled tasks (no manual supervision needed)
Data collection no longer requires me to manually trigger it every day. Through TRAE Work's scheduled task feature, I configured it to auto-execute every Monday at 7 AM:
Execute the full data-automation Skill workflow
Upload the merged result file to Feishu Cloud Docs
After configuration, on Monday mornings when I arrive at my desk, the latest data report is already generated.
Scenario 3: Feishu Cloud Docs auto-archive (more agile team collaboration)
In the past, after collecting data, I had to manually upload to Feishu, update document links, and notify team members. Now this step is automated too.
After the Skill finishes, TRAE Work will automatically call the Feishu CLI to upload the final Excel as a Feishu spreadsheet to Feishu Cloud Docs. Team members can directly view the latest data in Feishu, and on Monday morning they can simply open the data result link for a meeting review.
The entire flow changed from "manual web operations → manual Excel organization → manual Feishu upload" to:
"TRAE Work scheduled trigger → Skill auto-collection → auto cleaning/merging/validation → auto upload to Feishu spreadsheet → team directly views in Feishu"
*Sensitive data information in the image has been blocked / modified
Scenario 4: Data validation and quality assurance (preventing errors from flowing to the next stage)
This is the biggest difference from the previous workflow. Previously with manual operations, missing a single filter condition might not be discovered until days later, but now every step has validation:
Each core collection point must fully succeed
Each core collection point must have 30 rows of data
Unit conversion must兼顾 Chinese readability, numerical conversion and sorting must both be supported
The final Excel must contain both a Brand KOL detail sheet and a brand overview sheet
If any validation fails, the process stops immediately and saves a debug screenshot
How to Practice
I didn't start by having AI "directly write a complete script." Instead, I broke the work into multiple steps that can be verified by eye.
Step 1: First let TRAE Work understand the real process
I handed over my existing RPA flowcharts, Python scripts, operation recordings, and field requirements to TRAE Work, letting it organize the process first rather than rush to write code.
*Sensitive data information has been blocked / modified
Step 2: Execute in segments, verify each segment
The entire project advanced in phases: login verification → brand overview → full overview → filter verification → single KOL → full KOL → cleaning test → merging test → complete flow → Feishu upload → scheduled task configuration.
The most valuable part of this process is that TRAE Work can continuously read execution logs, screenshots, and Excel results, then modify the script — instead of explaining the project from scratch every time.
Step 3: Upgrade from "can run" to "can be scheduled"
Once the script stabilized, I had TRAE Work do three things for me:
Encapsulate as a Skill: supplement SKILL.md, configuration files, auto dependency installation, default output directory, result validation, so other environments can also call it;
Configure scheduled tasks: through TRAE Work's Schedule feature, set auto-execution every Monday;
Connect to Feishu: through the Feishu CLI, automatically upload results as Feishu spreadsheets.
How TRAE Work Solved the Hardest Problems
Challenge 1: Chrome native HID popup cannot be controlled via web selectors
After login, you need to click the avatar in the upper right corner, click "Update," and then select the USBKey from the Chrome native WebHID dialog.
Initially, I could only use screen coordinate clicking, but this is very sensitive to resolution and system scaling, making it unsuitable for packaging as a portable Skill.
Final solution:
Enumerate the VID/PID of the USBKey;
Write the Chrome WebHidAllowDevicesForUrls policy for the target website;
Default to disabling coordinate clicking, only keeping mouse clicking as a fallback when the policy cannot take effect.
Challenge 2: "Yi" (100 million) unit loss causing ranking errors
1.26亿 was truncated to 1.26, 3.21亿 was truncated to 3.21, and data in the hundred-million range ended up sorting after data in the ten-thousand range.
TRAE Work helped me trace back from the final Excel to the collection source, discovering that the sales amount split regex only recognized w, not the Chinese 万/亿. The final approach: preserve 万、亿、w completely during collection, keep original units in display values, uniformly convert to numeric values during sorting, and proactively search for疑似 truncated bare numbers during validation.
*Sensitive data information has been blocked / modified
Final Results
The final result is a complete automation solution of Skill + scheduled task + Feishu Cloud Docs:
Three-layer automation architecture:
Layer
Tool
Responsibility
Execution layer
Skill (data-automation)
Data collection, cleaning, merging, validation
Scheduling layer
TRAE Work scheduled task
Auto-triggered every Monday at 07:00
Collaboration layer
Feishu CLI + Cloud Docs
Results auto-uploaded as Feishu spreadsheets
Efficiency Improvement Comparison
Time comparison
Stage
Manual operation
After automation
Overview collection
~60 minutes
AI executes while I sleep
Detail collection
~1.5 hours
AI executes during my subway commute
Data cleaning and merging
~30 minutes
Feishu Cloud Docs upload
~2 minutes
Total
~3+ hours with manual supervision
AI executes without my involvement + unattended
Quality improvement
Previously with manual operations, issues like missing filter conditions, pagination omissions, and unit loss occurred frequently, often discovered during meetings requiring data rework.
Now: AI executes, I simply submit results
Changes in work style
Before: Every Monday morning manually open the data platform → search one by one → switch pages → copy fields → paginate → clean → merge → upload to Feishu → notify the team
Now: Every Monday morning open Feishu, and the data is already there .
5 TRAE Work Usage Tips
Tip 1: First let AI restate the process, then have it write code
If AI doesn't even understand the real operation path, the faster it writes, the more rework there will be.
Tip 2: Turn "problems you can see with your eyes" into executable validations
Don't just say "the promote button seems not to have been clicked," but instead: "first determine the position of the 'Promote' button, then attempt the operation; if the Promote button is still not detected, stop collection and save a screenshot." The problem won't quietly recur in the next execution.
Tip 3: After a single core collection point passes validation, then run the full pipeline
Full execution takes nearly 50 minutes. Without first validating individual core collection points, any small error will waste a whole run.
Tip 4: Have AI check the final file, not just check "the script"
The real completion standard is not the terminal showing success, but the Excel file existing, row count correct, fields not misplaced, monetary units not lost, and sorting consistent with business logic.
Tip 5: Learn to use Skill + scheduled task + Feishu to achieve "closed-loop automation"
After the script stabilizes, upgrade your workflow in three steps
Encapsulate as a Skill: so AI can call it directly
Configure scheduled tasks: so execution is schedulable
Connect to Feishu Cloud Docs: so results are collaboratable
This step transforms the project from "runs on my computer" to "data automatically appears in the team's Feishu every Monday morning."
Final Thoughts
What impressed me most about this practice is: the truly valuable aspect of AI programming is not just helping me write a few hundred more lines of code, but helping me gradually transform a vague, error-prone, experience-dependent office process into a system that is executable, verifiable, schedulable, and collaboratable .
From the initial "how to click a certain button," to "how to let other AIs also execute the Skill," to "how to make data automatically appear in Feishu every Monday": the project's goals kept upgrading.
The final delivery is not just an Excel, but a set of unattended production lines that can stably produce data and automatically archive it into a team collaboration space.