Recorded session · one instruction · 5 posts · 6m 19s
Qwen Work reads one sentence and hands the browsing to Tabbit. Tabbit opens the explore page, searches the keyword, ranks the results by likes, opens the top five one by one, and reads the copy and images. The agent then works out what those posts have in common and writes a file: sample breakdown, three title options, full body copy and a six-image storyboard.
使用 Tabbit 打开 https://www.xiaohongshu.com/explore，然后帮我搜索千问办公，在搜索结果里找到最热门的五个帖子，并点击查看详细内容和图片。然后模仿这些爆款，帮我写一份千问办公的帖子的内容脚本，包括标题，正文和图片脚本。
One paragraph, no code. Everything after the first clause is plain language.
00 · Full recording
67 seconds, unedited. Qwen Work plans on the right, Tabbit browses on the left. Every number on this page comes from this one session.
01 — Who decides, who executes
Nothing here is a hard-coded scraper. One side decides what to do; the other actually does it in a real browser.
Decision side · Qwen Work
Execution side · Tabbit
02 — Execution path
Timestamps come from the recording.
One paragraph carries the site, the keyword, the “top five by likes” rule, the request to read copy and images, and the output format.
Tabbit opens the explore page and searches the keyword. The log reports each action as it happens.
The agent sorts the results by likes and lists the top five with their counts: 1040, 819, 209, 180, 151.
Each post is opened in turn — copy, tags and images are collected, one post every 6–10 seconds.
The first post came back with the wrong title because the selector matched the wrong node, so the agent re-extracted it precisely. Later the card order shifted, so it switched to opening a post by its note ID.
The agent states that all five posts have been read, then writes the script file: sample table, shared pattern, three titles, body copy, six-image storyboard, publishing notes.
03 — What it collected
The ranking rule was “most liked”. The counts below are the ones visible on screen during the run, not live values.
「千问办公」打通班群，成绩通知全自动！
桌宠教程来了！
博士锐评AI办公排名？
Vibecoding一个A股选股工作台
千问办公国际版，开启公测
Same five posts, with the title text as it appeared in the search results.
| Rank | Post title | Likes | Format |
|---|---|---|---|
| 1 | 「千问办公」打通班群，成绩通知全自动！ | 1,040 | Video post |
| 2 | 桌宠教程来了！ | 819 | Video post |
| 3 | 博士锐评AI办公排名？ | 209 | Image post |
| 4 | Vibecoding一个A股选股工作台 | 180 | Video post |
| 5 | 千问办公国际版，开启公测 | 151 | Image post |
Titles, counts and log lines are read from the recording and are shown here only to explain the workflow. This page does not host or republish any third-party post.
04 — What it worked out
These lines are from the agent’s own analysis, written before it drafted anything.
标题=身份/场景+惊叹式成果+标点收尾
Title = identity or scene + a result worth exclaiming about + punctuation
开头第一句先讲“我被什么折磨过”
The first line says what was making the author miserable
正文只讲3步，强调“我没写一行代码”
The body lists three steps and stresses “I did not write a single line of code”
结尾有彩蛋：存成 Skill 下次一键复用
The ending has a payoff: save it as a skill and reuse it next time
标签=身份标签+场景标签+工具标签
Tags = identity + scene + tool
封面图大字标题，内页图=步骤截图+成果展示
The cover carries the headline in large type; inner images are step screenshots plus the result
The formula it wrote down
真实人设 + 具体痛点 + 三步极简流程 + 成果对比 + 提示词放送 + 场景化标签
05 — What it delivered
The output of the run is one Markdown file with four sections. Below is what is in it, quoted from the recording.
Generated file
Four sections in the file
Three title options it wrote
The body above is the copy the agent wrote in the recording. It is the example output of the run, not a template you have to reuse.
06 — The six-image storyboard
The file does not stop at copy. It specifies six frames — what is on each one and how it is laid out. The composition notes below are quoted from the generated file.
From the file米白/浅灰底，第一行身份、第二行“搞定一件事”，“10 分钟”用红色高亮
From the file左“以前的我”凌乱桌面拼贴配“翻记录翻到眼花”，右“现在的我”一句指令，底部横条“从 3 小时 → 10 分钟”
From the file三张横向步骤卡 1/2/3：上传资料 → 输入指令 → 确认发送，配截图与红框标注
From the file生成结果全文截图（脱敏），红笔旁注“这段数据它自己从文档里抓的”
From the file清单式卡片“它还能替你干这些”：会议纪要 / 邮件措辞 / 月度汇报 / 存成 Skill
From the file“你每周写周报要多久？评论区报个数”+ Skill 保存入口截图
07 — Why this holds up
None of them depend on how large the model is.
01
No API key, no SDK, no scraper to maintain. If a person can read it on the page, the agent can read it — including content that only appears after a click.
02
Most setups stop at saving JSON or Excel and leave the analysis to you. Here the same session that read the posts also wrote the breakdown and the draft.
03
What comes back is a file an editor can act on: a comparison table, a formula, three titles, body copy and six frames with composition notes.
08 — Replicate it
Paste these into any agent that can call Tabbit.
01
Naming the page and the sort order is what keeps the shortlist stable from run to run.
Use Tabbit to open https://www.xiaohongshu.com/explore, search for [keyword], and take the five most-liked posts.
02
A rewrite alone loses the research. A table forces the agent to say why each post worked.
Open the file with a table comparing the five posts: title pattern, core pain, audience, and the point that drives interaction.
03
On image-heavy platforms the cover and the first three images carry most of the weight. Ask for composition, not “some pictures”.
For the image script, plan six frames: frame 1 cover composition and headline; frames 2–5 steps and result comparisons; frame 6 the closing question and where to get the material.
04
Say what the file must contain up front and you skip the follow-up round trip.
Save the result as a Markdown file with: sample breakdown table, shared pattern, three title options, full body copy, six-image storyboard, and publishing notes.
09 — Keep reading
Same pattern, five different jobs: a general agent does the thinking, Tabbit does the browsing.
Doubao Work plans, Tabbit reads Eastmoney, the report renders itself.
Qwen Work plans, Tabbit reads Xiaohongshu, the script file is the deliverable. You are reading this one.
Codex reads the spreadsheet and images, Tabbit fills the merchant console, the storefront is checked.
DeepSeek designs the schema; Tabbit writes 200 orders from a CSV into a Feishu Base.
WorkBuddy filters and writes; Tabbit visits four selected sources.
Benchmark report
The same tasks run across Tabbit, Codex Chrome and Agent Browser, scored on answers, median time and input tokens per correct answer.
FAQ
No. The whole run is one paragraph of plain language. There is no script, no SDK and no API key — the agent drives a browser the way you would.
Yes. In the recording the agent reads the copy and the tags, and opens the images of each post. The image script it writes afterwards is based on what those posts actually did.
It recovers instead of stopping. Twice in this run: a title selector matched the wrong node, so the agent re-extracted the title precisely; later the card order shifted, so it opened the post by its note ID instead.
It is a draft you can edit, not a finished post. The value is that the research, the pattern and the structure are already in one file, so editing is the only step left.
The status capsule in the recording reads 6m 19s for five posts, including two corrections and the final write. The 67-second video is a speed-up of that session.
Get started
Download Tabbit, hand it to the agent you already use, and point it at the page you would otherwise open twenty times.
Free. macOS and Windows.
Tabbit