Recorded run · one instruction · 4 sources · 4m 22s
WorkBuddy breaks down the job and calls Tabbit CLI to operate a real browser. It checks each source, reads the pages, keeps the relevant AI developments, removes overlap and delivers a structured Chinese-language brief.
@tabbit 打开并阅读 smol.ai、Hugging Face Daily Papers、AIHot 日报和橘鸦 AI 日报，收集今天值得关注的 AI 信息，去重、分类并整理成一份中文日报。
@tabbit open and read smol.ai, Hugging Face Daily Papers, AIHot and AIya Daily. Collect today’s important AI developments, deduplicate and classify them, then write a Chinese daily brief.
@tabbit hands browser execution to Tabbit. The rest is ordinary language.
00 · The recording
No cuts, sped up from the real session. WorkBuddy plans and filters on one side while Tabbit opens four sources in a real browser on the other. Every step below comes from this session.
01 — Roles
This is not a fixed scraper. The agent responds to the goal, each site’s state and changes in the visible page.
Decision side · WorkBuddy
Execution side · Tabbit CLI
02 — Workflow
The recorded run can be reduced to five steps.
WorkBuddy confirms four sources and the requested output.
Tabbit opens smol.ai. With no new issue that day, the task reads the latest archive instead.
Tabbit visits Hugging Face Daily Papers and reads titles, institutions and research directions.
AIHot and AIya contribute product updates, local industry news and Chinese coverage.
WorkBuddy merges the sources, prioritises the items and produces the daily AI report.
03 — Browser evidence
These frames come from the same 45-second recording and show Tabbit moving through the sources.
When smol.ai has no new issue, the task enters the archive and reads the latest edition.
Daily Papers contains many entries; the agent keeps the research directions useful for the brief.
AIHot and AIya complement international news and papers with product and industry coverage.
04 — Deliverable
WorkBuddy combines all four sources so the reader does not have to copy, classify and deduplicate the items again.
The events most worth immediate attention.
Model, framework, speech and multimodal updates.
Tools, cloud services, compute and partnerships.
Institutions, directions and findings from the daily paper list.
This page documents one recorded workflow. Source content changes by date, and execution time varies with page and network conditions.
05 — Why it works
Checking freshness, selecting items, reconciling sources and structuring the output take longer than opening four tabs.
A fixed source list gives the task clear boundaries and leaves pages that can be checked.
A missing daily edition or changed layout becomes a decision in the run, not an immediate failure.
Name the sections in the instruction and the task ends with a deliverable instead of raw notes.
06 — Reproduce it
The important part is making the rules explicit.
01
This reduces search drift and makes citations easier to review.
02
For example: use the latest issue and label its date.
03
Name the industry, region, company set or themes to keep unrelated news out.
04
Ask for headlines, models, industry and papers to avoid a second editing pass.
07 — More cases
Five tasks, one pattern: the agent thinks and Tabbit operates the browser.
Doubao Work plans; Tabbit reads Eastmoney; charts and conclusions follow.
Qwen Work analyses top posts while Tabbit opens each post and its images.
Codex prepares records; Tabbit completes browser forms.
DeepSeek structures fields; Tabbit transfers web data into the table.
WorkBuddy filters and writes; Tabbit visits four selected sources. You are reading it.
Browser-agent benchmark
The benchmark compares Tabbit CLI, Codex Chrome and Agent Browser across 25 tasks and 225 runs.
08 — FAQ
No. WorkBuddy plans from the natural-language goal, then drives a real browser through Tabbit CLI.
Yes. Put the sites, filtering rules, missing-update behaviour and output format in the instruction.
Named sources keep the boundary clear and retain reviewable pages; cross-source deduplication also suits a recurring brief.
No. That is one recorded run. Timing changes with page loading, item count and network conditions.
Give your agent a real browser
Download Tabbit and type /tabbit in an agent that supports tools to begin a visible, reviewable browser workflow.
Available for macOS and Windows.
Tabbit