WorkBuddy Skills, Explained by the Permissions They Ask For

WorkBuddy Skills are packaged scripts, not prompts. That one distinction explains most of the risk discussed on this page: a Skill runs code with your identity and your file permissions, so the useful question is never what it does, but what it can reach.

01 /What are WorkBuddy Skills, precisely?

Tencent's documentation defines the feature by what it contains rather than by what it feels like. A Skill packages executable scripts and workflows so the assistant can complete specific actions under your authorisation. Three parts of that sentence carry weight, and most of the confusion around the feature comes from quietly dropping one of them.

The first part is that a Skill is code. A prompt asks a model to attempt something; a script either runs or raises an error. That is why a Skill can rename four hundred files in a folder while a chat instruction more often produces a description of renaming four hundred files. It also means the failure modes are ordinary software failure modes: a wrong path, a missing dependency, an API that changed shape last month.

The second part is the phrase under your authorisation. A Skill does not execute as a guest with a blank account. It inherits the identity the client is already using, which is exactly why the official liability notice reads the way it does. Whatever a Skill does to a file, it does as you. What the client itself is is worth reading first if the permission model still feels abstract.

The third part is specific actions. The normal shape of a competent Skill is narrow: one input format, one output, a handful of deterministic steps. Packages that claim to cover an entire department's workflow tend to be prompt libraries wearing a package wrapper, and they behave like it when the input is malformed.

02 /How do WorkBuddy Skills differ from Experts and Connectors?

All three entries sit in the same sidebar menu, labelled Experts, Skills and Connectors, and they are introduced together, which is why people mix them up. The distinction is about what each one actually changes.

A working rule: Experts change how it thinks, Skills change what it can do, Connectors change what it can reach. The rule is crude but it survives contact with the menu.

The three also differ in where the permission lives. A Connector's permission surface is the external service's own authorisation screen, which most people have used before and can revoke. A Skill's permission surface is your file system, your shell and sometimes your screen. Per Tencent's documentation, reading and writing through a connector does not itself consume credits, while browsing, uploading and managing skills is also free of charge. Consumption starts when the client executes a task.

03 /What permissions can a Skill request?

This is the section most write-ups skip, and it is the only one that matters before you install something you did not write. Per Tencent's own security notice, a Skill may request any of the following.

1. Account identity information.
2. Contact details.
3. Your input content.
4. Usage behaviour logs.
5. Third-party platform credentials.
6. Local file read and write.
7. Screen and input control.

The last two are the pair to think about. A package that can read and write local files is a package that can delete them; a package that controls the screen and the input devices is a package that can click on your behalf. Both together, running under your identity, is not a toy and should not be installed on a machine that holds work you cannot restore from backup.

Notice what is absent from that list: there is no permission called "safe subset". A Skill does not get a sandboxed guest account by default, and the client cannot tell you in advance which of the seven a given package will use. You find out by reading the package.

04 /Where does the real risk sit?

Tencent's documentation is unusually direct on this point, and it is worth quoting rather than paraphrasing.

Verification noteNon-official Skills may carry malicious prompt injection, privilege escalation, or backdoors. Before installing a third-party Skill, carefully review its permission requests, source documentation, and script contents, and confirm the scope of data being sent externally.

And on who carries the consequence:

When a Skill calls a third party, modifies local files, or executes system commands, it does so as you. For file deletion, bulk writes, system configuration changes, or any financial operation, verify on a small scale before scaling up.

Two practical consequences follow. Read the script. A Skill package is code that will run on your machine, so if you cannot read what it does, treat it as untrusted regardless of how many people recommended it. Then establish whether data leaves the machine. Some Skills send your input to a third-party service as a design decision rather than a flaw, but you want to know that before you paste a client's spreadsheet into one, not afterwards.

There is also a supply-chain angle that has nothing to do with malice. A Skill that shells out to a CLI tool inherits that tool's version, and a Skill written against last quarter's API may fail quietly rather than loudly. Silent failure is worse than a crash when the output is a file someone will act on.

05 /What do WorkBuddy Skills cost to run?

Skills consume credits, and complex ones consume more. Tencent flags this explicitly: tasks involving complex work and complex Skills usually consume more credits. The official pages do not publish a per-call number, and this site will not invent one, so treat any article quoting an exact figure per Skill invocation as unverified until it links to a source that publishes it.

The behaviour that matters is the shape of the cost, not the number. A Skill that loops over a hundred files is a hundred small operations, not one prompt. That is the same reason the security advice says to test on a small batch first: running a package on five files is both the cheap way to check the credit cost and the safe way to check what it does. It is also the reason WorkBuddy Skills that run unattended on a schedule deserve a bounded input, rather than a folder that grows every week.

06 /Which WorkBuddy Skills are listed on this page?

Four cards below exist because people search for those four terms. None of them has been run on this machine, so each card says so instead of describing a feature it might have. A card here is a search entry, not a recommendation.

The rule that keeps this list short is deliberate: a Skill earns a page after a run, not after a README. Ranking lists of fifteen Skills are cheap to generate and mostly worthless, because they are written from package descriptions rather than from use. A page on this site is supposed to include the parts that only appear in practice, such as which permission the package actually requested, how long a real run took, and what it did when the input was malformed. If a Skill you need has no page yet, that is usually the reason, and the three routes into the client are documented separately from the review itself.

Five packages from this machine were cleaned and published on 2026-11-11, each with its own page explaining why it was written, how it was built and who should use it: the SEO Dev Spec numbered rulebook, the GEO AI Traffic probing Skill, the presale validation framework, the game site development method and the game site selection method. The five public repositories are seo-dev-spec, geo-ai-traffic, presale-validation, gamesite and gaming-site-selection. Their rule sources are in Chinese and their READMEs are English guides, not translations.

What is verified on this page and what is not, stated plainly: the permission list, the two quotations and the credit behaviour were checked on 2026-11-10 against Tencent's own documentation and security notice. The four cards above were not tested, no Skill package was executed to write this page, and no claim is made here about how any of those four behaves in practice. That distinction is the whole editorial rule of this site, and it is cheaper to state than to imply.

07 /How does a Skill earn a page here?

The short answer is that it gets run twice. Once on a sample that should work, and once on an input that is deliberately slightly wrong, because the second run is where a package reveals whether it stops or continues on bad data. A Skill that continues on bad data has produced files you now have to audit by hand, which is more expensive than the task it was supposed to save.

The longer answer is a checklist held to every package, and the same checklist applies to the custom work described elsewhere on this site. WorkBuddy Skills written to order do not get a different standard from published ones, which is why the custom page asks for a sample of the correct output before anyone quotes anything.

None of those five checks requires reading a single line of marketing copy, which is the point. If you are evaluating a package yourself and cannot answer the failure behaviour row, you have not finished evaluating it.

08 /Frequently asked questions

Do WorkBuddy Skills run on my machine or in the cloud?

Both are possible, and the distinction is not decorative. Per Tencent's cloud documentation, cloud tasks run in an isolated sandbox and do not support the local workspace, full access or custom models, while scripts in local tasks execute on your own machine. WorkBuddy Skills that touch local files therefore depend on the local execution path, and moving the same task to the cloud changes what it can reach.

Are third-party WorkBuddy Skills safe to install?

Not by default, and the vendor does not claim they are. The official notice warns about malicious prompt injection, privilege escalation and backdoors in non-official packages, and states that when a Skill calls a third party, modifies local files or executes system commands, it does so as you. Safety here is a reading exercise plus a small-scale test, not a setting.

Can I disable a Skill without uninstalling it?

Yes. Installed Skills can be disabled without removal, and Tencent recommends enabling only what the current task needs. The recommendation is about more than tidiness: every enabled Skill is another candidate for invocation on the wrong task, and Skill calls consume credits. Disabling is reversible and takes one click, so there is no reason to keep a rarely used package enabled. If you accumulate WorkBuddy Skills the way people accumulate browser extensions, the keyword box above the list is the only thing that will still work a year from now.

Do WorkBuddy Skills work on Linux?

The client is distributed for Linux, but the official download page only points UOS and Kylin users at their system app stores, and the packages for other distributions are reachable through an update endpoint that is not linked from the page. Skill behaviour also depends on what the package shells out to, so a package that assumes macOS tooling will not become portable just because the client installed. The verification steps for the Linux packages are documented separately, including the two SHA-256 values and the eight fields the endpoint returns.

09 /Where should you go next?

• Need to get one into the client? Three install paths compared
• Need one that does not exist yet? Custom Skill work and how to ask
• Not installed yet? WorkBuddy download, every platform
• Want the dated field records first? Notes from hands-on testing
• Running a whole pipeline? WorkBuddy automation with Skills, MCP and schedules