跪拜 Guibai
← Back to the summary

Ten Unorthodox WorkBuddy Moves: Persistent-Memory Cron, Session Hijacking, and One-Click Replication

10 Unorthodox Uses in WorkBuddy

As WorkBuddy keeps updating and iterating, it gains more and more features.

I've been using WorkBuddy for a while now. Taking advantage of the last couple of days, I've summarized 10 very practical features in WorkBuddy.

You could also call them unorthodox uses, hehe.

01 Automation

Many friends understand automation as just a scheduled task. In essence, that's not wrong, because automation in WorkBuddy is indeed a scheduled task. But what I want to say is that it's actually cron with persistent memory, and scheduled tasks that can bind experts, skills, and working directories.

Combining these three mechanisms can make scheduled tasks take on a wide variety of effects.

The simplest examples are the daily work report and the project health report.

Daily Work Log

Summarize the day's work and generate a work log that can be **shared directly with others**.

Data sources (try all three; missing any one does not prevent the output):
1. Retrieve all conversation records with me for the day, categorized into several independent work items by topic.
2. Check project directories with file changes that day (business directories in the current workspace, excluding dependency packages and cache directories), as evidence of "actual output".
3. If there is already a project-facing inspection report or daily report for that day, merge its conclusions.

Output: markdown file, written to worklog-YYYY-MM-DD.md in the current working directory (date according to system local time).
Fixed structure:
# Work Log YYYY-MM-DD
## Main Direction of the Day (within 3 lines, only the direction, no details)
## Work Item Progress (for each item: goal / what was done / conclusion / blockers / next steps)
## Unclosed Items (points awaiting confirmation or not yet verified, listed one by one)
## Tomorrow's Priorities (at most 3)
## Reusable Gains (practices, scripts, checklists formed this time; omit this section if none)

Redaction rules (highest priority, sacrifice specificity for privacy when in conflict):
- No real names, company names, client names, phone numbers, emails, accounts, domain names, IPs, ports, intranet addresses, absolute paths, keys, or tokens.
- Paths only retain relative positions, prefixed with "workspace", e.g., "workspace/project directory/report module".
- People are always written as "collaborator", clients as "Client A / Client B", internal systems as "internal system", third-party services as "a certain external service".
- Do not quote session details or code snippets; only write summarized conclusions. All proprietary identifiers are generalized.
- Do not guess unconfirmed information just to be specific.

Other rules:
- Only write content with evidence; if not found, write "no record". Do not fabricate.
- If the target file already exists, append a "supplementary note" section; do not overwrite the original content.
- If there are no conversation records and no file changes that day, only reply NOOP and do not create a file.
- After completion, reply with a summary within 3 lines; do not repeat the full text.

image.png

Here's a look at my work log. It's been running for a few days now and it's been very stable so far.

image.png

I set this automation to run once every night at 10 PM, summarizing the day's work so I can stay on top of things.

Moreover, the scheduled task does the summary within the same session, which makes it easy to see the log records, and it becomes convenient to build a workbench out of it.

(Actually, I wanted to make a workbench for the scheduled tasks, but maybe I didn't describe it clearly. WorkBuddy just made a workbench for all my scheduled tasks.)

image.png

There's also a project health daily report that shows the health status of your local services.

Project Health Daily Report

Check the health status of several local services in the current workspace and produce an anomaly-priority brief.

Check items:
1. Port liveness: Check whether key service ports have a process listening; if listening, request the health endpoint once, and record the returned content or failure reason.
2. Recent activity: For each project directory, list the 3 most recently modified files and times (excluding dependency packages and cache directories).
3. Dependency health: Whether the dependency manifest exists and its modification time is newer than the dependency directory; if there is a lock file but dependencies are not installed, mark as not installed.
4. Stale check: The date of the newest file in the output directory, and the last modification date of the status file. If either exceeds 3 days, mark as "stale".
5. For projects that are not git repositories, skip all git commands; do not attempt them.

Output: write health-YYYY-MM-DD.md (date according to system local time), divided into "abnormal / warning / normal" three levels.
In the chat, give only a summary within 5 lines, sorted by abnormal to warning to normal, and only write items needing action.
Rules: only report actual measured results; if a command fails, record the failure reason as is; do not speculate. When everything is normal, write the file as usual, but reply only NOOP in the chat.

For example, from the scheduled task execution in yesterday's log, there seems to be a problem now.

image.png

I also ran a workbench for this project health daily report.

image.png

As for the "cron with persistent memory + binding experts, skills, and working directories" mentioned at the beginning, opening WorkBuddy's "New Scheduled Task" configuration dialog makes it very intuitive:

image.png

02 Skills

If I had to pick the one feature with the most long-term value, I'd choose Skill.

A Skill is fixing a set of best practices you've figured out yourself into a workflow. When you encounter a similar task again, you just follow the Skill's operation steps.

In WorkBuddy, many people easily confuse Skill with Expert. I did too at first, but Skill and Expert are different.

Here's the simplest distinction:

Dimension Skill Expert Expert Team
Nature Tool capability AI consultant AI collaboration team
Composition Scripts + workflow Persona + methodology + toolchain Multiple experts + collaboration process
When to choose When AI needs a certain capability A clear single-point problem Complex tasks requiring multi-role collaboration

In one sentence: Skill is capability; Expert is capability + experience; Expert Team is multiple experts + collaboration process.

A Skill directory looks like this:

my-skill/
├── SKILL.md            # Required, the skill's manual
├── scripts/            # Optional, scripts that can run directly
├── references/         # Optional, reference documents read when needed
└── assets/             # Optional, templates, images, fonts used for output

The top part of SKILL.md is YAML frontmatter. name and description are required.

For example, let me dissect a Skill I use. Its real directory tree looks like this.

image.png

This is a third-party "AI Content Inspector" installed from the Skill Market. It doesn't produce content; it only scores existing content, finds flaws, and repeatedly rewrites it until the "expert panel" gives a score above 90.

It runs this process according to the content of Skill.md.

image.png

In the entire skill, description is the most critical field in the whole mechanism. WorkBuddy uses it to decide whether to load this skill for a conversation. If you write it vaguely, the skill won't be triggered at all.

For example, this Skill called Content Ops has an empty string "" for description in its SKILL.md. So unless you explicitly @ it, it won't trigger.

image.png

My judgment: when you're about to do the same thing a third time, you should make it a Skill.

Here's a template for creating a Skill.

Package the process I just described into a Skill and place it in the user-level directory.
Requirements: write a clear description of the trigger scenario; extract repeatedly used commands into scripts under scripts/; put output templates in assets; after writing, verify with a new example that it can be triggered correctly.

It's a pity that after writing a Skill, it's easy to just throw it away. I suggest reviewing them periodically and either fixing the description or deleting any that have never been triggered.

03 Browser Control

Many AIs can access the internet, but internet search and directly controlling your browser are two completely different things.

Ordinary internet access tools or headless browsers can only fetch public web pages. They get stuck when they encounter a backend that requires login, a dynamically loaded list, or a detail page that only appears after clicking a popup.

WorkBuddy's browser control, however, connects directly to your Chrome / Edge / QQ browser via a browser extension, which means it can directly borrow your login state, cookies, and currently open tabs.

Enabling it is simple: go to "Settings → Agent → Browser Control", turn on the switch, and install the corresponding browser extension:

image.png

Everyone can think of the usual uses: help me fill in a form, run a webpage regression test, take a screenshot. But what really makes this feature shine are these unorthodox uses:

  1. Turn consoles without export functionality into exportable data sources

Many SaaS backends, order lists, financial reconciliation pages, or internal data dashboards don't provide APIs, and some don't even have a CSV export button. Previously, you had to flip through pages one by one.

Now you just log in and open the list page in your browser, then let WorkBuddy handle the rest: automatically paginate, click each row's detail popup to fill in hidden fields, and compile everything into a structured CSV saved to the workspace. It's like lending it a pair of eyes and hands to turn any web backend into a local data source for free.

Look at the backend order page currently open in my browser, keep the current login state, don't log out. Help me flip through the past 30 days of records page by page, extract the "order number, amount, status, order time" and compile them into a CSV saved to the current workspace; if there are fields that only appear when clicking "View Details", click them to fill in the missing information. Note: do not click any buttons related to refunds, deletions, or submissions.
  1. Conversely, "read instead of write": feed a local draft into a multi-step web form

When doing content distribution, product listing, or filling out various tedious backends, the most annoying part is that you've already organized all the information locally in Markdown or Excel, but in the web backend you have to break it into a dozen input boxes and paste them one by one, select dropdowns, and upload fields.

The unorthodox approach is to hand the local file path and the logged-in backend page to WorkBuddy, have it fill in each field according to the local content, even upload corresponding attachments automatically, and finally save a draft and screenshot it for your review before publishing.

Read draft.md in the current directory. According to its title, body, summary, and tags, fill them one by one into the backend editing page currently open in my browser. Strictly: only fill in content and click "Save Draft"; absolutely do not click "Publish". After completion, provide a screenshot for me to verify.
  1. Use it for no-code-required, logged-in backend patrols

Some competitor updates, data dashboards, or backend alert pages can only be viewed after logging in. Writing Playwright scripts yourself is a hassle just for maintaining cookie expiration. Now you can borrow the browser's existing login state, have WorkBuddy periodically check specified tabs, diff against the last locally stored record, and send you a summary when there's a change.

04 Human-AI Co-Writing

I really like this feature because it solves the most awkward part of AI writing.

The old process: you ask AI to write a section → copy it → paste into Word → adjust formatting → discover the third sentence is wrong → go back and ask again → copy again.

image.png

Human-AI co-writing means: directly select a section in the document and have it modify it in place.

image.png

It can modify Word, Excel, PPT, and Markdown. Select and modify, and see the result immediately after the change, without having to save and reopen to check the effect.

Change the selected section to the style for external release: keep the original meaning, remove internal jargon and colloquial expressions, and don't modify any sentence outside this section.

I deliberately added "don't modify any sentence outside this section" in that prompt. In such delivery scenarios, the biggest fear is AI making random changes.

Clear scenarios where it fits:

Scenario Why it fits
Commercial orders, finalizing manuscripts Keep the original style, only change local parts
Adding a formula column to a table No need to rebuild the whole table
Changing title or data on a PPT slide Only change one page, touch nothing else
Unifying terminology in long documents Change piece by piece, see results immediately

05 Table Analysis and Visualization

If you already have an Excel or CSV file, maybe an exported order list, a questionnaire result, or a monthly breakdown from finance, you don't need AI to create a new table for you; you need it to understand the table.

Read this table, first tell me: how many rows, which columns have missing values, and whether the date columns have a consistent format.
Then do a cross-tabulation of "month × category", identify the top three items with the largest month-over-month change, and draw a chart.
All conclusions must be based only on the data in the table; if something cannot be calculated, don't guess, list it separately as "needs supplementary data".

Things it can do:

06 Connectors + Custom MCP

What do connectors do? In one sentence: they are the bridge between WorkBuddy and the outside world.

With just a large model, an agent can only use information you give it. With connectors, it can actually read your emails, browse your documents, check project boards, and send meeting invitations.

By use, they roughly do four things:

Scenario Description Example
Data query Fetch information from external data sources Query databases, retrieve knowledge bases
Service invocation Call third-party APIs to complete operations Send emails, create schedules, create tasks
File management Access cloud storage and file systems Read cloud drive files, upload attachments
Message notifications Integrate with IM tools Send WeChat Work / Feishu messages

But I think the real dividing line is custom MCP.

No matter how many existing services an agent supports, it still depends on what its ecosystem offers. With custom MCP, the question becomes what you want it to support.

Many vendors stop after integrating the top applications. But many people want something more personal: local note bases, albums on a NAS, internal dashboards, or a small API you wrote yourself. These aren't things the official team will do for you, but you can write an MCP to hang them on.

Add a custom MCP service to the configuration file. Only merge new entries, don't overwrite existing config. After writing, show me the newly added section for verification; don't start the service yet.

For example, I mounted my Typora note library as an MCP, so I can read article content from outside via MCP.

Typora note library mounted as MCP, actual test

Also, a reminder: MCPs newly added in WorkBuddy don't take effect automatically. You need to go to the "Custom Connector" entry at the top right of the connector management page and click "Trust" to enable it.

image.png

07 Data Library

The data library is the most underestimated feature among these. Many people think it's just a network drive.

Actually, a more accurate definition is: the shared drive between you and the agent.

Documents, tables, and HTML pages generated by tasks get collected here, and files you upload yourself are also here. Human and agent outputs are managed in the same place. You can see space usage in the lower-left corner.

There's an introduction to the data library in WorkBuddy that I think is well written; take a look.

image.png

Its real value comes in three layers:

  1. Basic: storing and retrieving materials.

  2. Intermediate: human-AI co-writing. Multiple people collaborate on the same document, AI revisions, and paragraph comments.

  3. Advanced: build a personal workbench with the data library. HTML + CSV can be made into a lightweight app, turning local files into an online application.

    Store this task's outputs into the data library, and create a directory by project for classification. Then publish that HTML as an online link and give me the link.

The third approach is the most valuable. Many vibe-coded things end up being usable only by yourself, and sharing them with others is troublesome. Once it's published as a link, it becomes something others can actually click and open.

By the way, the data library and the knowledge base are not the same: the data library manages outputs; the knowledge base manages content for long-term accumulation and repeated retrieval.

08 Experts / Expert Teams

I mentioned Experts briefly when discussing Skills. Let me expand on it here, because it's indeed the most distinctive design in WorkBuddy.

An Expert is not just a persona that changes how it speaks; it's persona + methodology + toolchain. A financial expert and a legal expert, facing the same material, will differ in what to look for, how to judge, and how to format the deliverable.

An Expert's profile describes its capabilities, areas of expertise, and task examples. Take this Expert, for example.

If you're a beginner wanting to build a WeChat mini-program and don't know how, you could directly ask the AI, but with an Expert's assistance, it lowers the entry barrier and saves you the tokens you'd spend stepping on pitfalls.

image.png|437

An Expert Team, on the other hand, groups multiple experts into a collaborative team. For example, an MVP development expert team might include a project director, product manager, UI designer, DevOps, and frontend tester, essentially replacing the roles of a human team with agent members.

image.png|451

Take this MVP development expert team; it includes many roles. These roles follow different capability responsibilities to complete the specific parts of a project you need. It's like hiring a small company to assign tasks, and all you pay is some tokens.

But there's a cost issue that must be stated clearly: expert teams call multiple experts in parallel and involve multiple rounds of model interactions; credit consumption is usually 3–5 times that of a single expert.

My usage is generally conservative:

Situation What to choose
A clear single-point problem Single Expert
Need to deliver a whole set, requiring multi-role collaboration Expert Team
Just need AI to have a fixed capability Skill, not Expert

Also, binding an Expert into an automation is an interesting combination: the same scheduled task, with a different Expert, is like different people showing up to work on time, haha. Except you can't dock their performance.

There are also so many Experts and Expert Teams now that their responsibilities often overlap. That's a problem too.

09 WeChat Assistant: Remote Task Assignment

This is what I use most when I'm traveling or away from my computer.

In the configuration, there's an "Assistant" entry. After binding, you can send commands to WorkBuddy directly from WeChat, letting it do the work on your computer.

image.png

This directly enables remote task assignment.

Here's a chain I've successfully run: periodically grab a batch of content → save to knowledge base → merge duplicate information and generate a "new information brief" → create an HTML page → push to WeChat.

image.png

One limitation to note: some content sources don't expose a callback interface for new articles, so you can't get notified the moment something is published; you have to poll periodically. That's not a WorkBuddy problem; the upstream just doesn't offer it.

10 Inspiration One-Click Replication

You must try the Inspiration feature in WorkBuddy's Inspiration Community.

image.png

This feature lets you replicate the same thing with one click. If you don't know what to build, you can come to this community and take a look.

The Inspiration Community solves the cold-start problem: it's a collection of real works that others have already successfully run, including various personal workbenches, data dashboards, interactive pages, and analysis templates.

After clicking one-click replication, you can directly use the finished effect of someone else's work.

image.png

My experience: replicate first, then fine-tune; it's almost always more efficient and consumes fewer tokens than starting from scratch with a Prompt. Even if half the pages need to be completely redone, having a reference point that runs gives you a head start and much lower trial-and-error costs.

I've listed 10 fairly common and practical features of WorkBuddy, but WorkBuddy has many more features; ten is not enough to cover them all. The following are ones I also use in daily work:

Finally, as usual, here's a quick-reference table:

Feature One line When to use
Automation cron with memory, can bind experts and skills For repetitive things, when something should happen at a scheduled time
Skills Turn a workflow into a capability When you're about to do the same thing for the third time
Browser Control Borrow your real login state to take over web pages For cross-page data scraping, auto-filling forms, backend patrols
Human-AI Co-Writing Select and modify in place, save in place When an existing document needs partial changes while keeping the overall style
Table Analysis & Visualization Understand the table you have When you have data and need conclusions and charts
Connectors + MCP Let the agent understand the external world When you need to read emails, browse documents, send messages
Data Library Shared drive for humans and agents For output storage, collaboration, publishing links
Experts / Expert Teams Capability + experience / multi-role collaboration Find an expert for a single-point problem; use an expert team for a large project
WeChat Assistant Assign tasks and receive results from WeChat When you're away from your computer but still need results
Inspiration One-Click Replication Make the same version from a proven great work When you don't know where to start or want to quickly customize a personal version