Skill packages and safety
Last updated:Nov 15, 2026
This page describes what an official yunyoo skill package looks like and what makes it trustworthy before and after release, for readers who want the details. There are only official skills at the moment and we do not accept third-party submissions yet; when that changes we will describe the process separately. For installing and using skills see Skills.
What is in a package
A package is a ZIP that contains a single top-level folder named after the skill:
| Path | Content |
|---|---|
SKILL.md | Required. A short front matter block (name and description) followed by the text written for the AI assistant |
yunyoo.skill.json | Required. The manifest: version, publisher, required permissions and tools, and so on |
LICENSE | Required |
references/ | Reference files the text points to when needed |
assets/ | Templates and images |
scripts/ | Scripts, each listed in the manifest. The first official skills ship no scripts |
Naming rules
- The name starts with
yunyoo-and uses only lowercase letters, digits and hyphens, for exampleyunyoo-app-smoke. - The folder name, the
nameinSKILL.mdand thenamein the manifest are the same. - The version has three parts, for example
1.0.0. Once a version is registered its content never changes; any change needs a new version. - The
descriptioninSKILL.mdis written in English, in the third person, says what the skill does and when to use it, includes Chinese trigger words, and is at most 1024 characters. The AI assistant uses this text to decide whether to use the skill.
Manifest fields
The manifest is the disclosure, for people and for programs: the desktop app shows it to you before installing, and the platform checks the skill text against it.
| Field | Meaning |
|---|---|
name, version | Name and version |
publisher | The publisher; only official for now |
displayName, summary | Display name and a one-line summary, each in English, Simplified Chinese and Traditional Chinese |
category, tags | Category and tags |
license | License |
interfaces | How the skill can be used: mcp, cli or both |
requires.cli, requires.mcpServer, requires.mcpProtocol, requires.desktop | Version requirements for the command line, the MCP server, the MCP protocol and the desktop app |
requires.mcpTools, requires.cliCommands | The MCP tools and command-line subcommands it uses |
requires.scopes | The token scopes it needs |
requires.permissions | The account permissions it needs |
requires.network | Hosts it contacts: the platform hosts its tools and commands talk to, and the hosts of any links in the text |
agents, nodes | The AI assistants it was verified with; whether it applies to local nodes, cloud nodes or both |
risk | The highest risk level: read-only, writes, destructive, spends money or security related |
scripts | The scripts in the package; empty when there are none |
Writing for both ways of use
A skill can have the assistant use the MCP tools, the command line, or either. Both must be declared in the manifest, and the text always writes them like this:
- an MCP tool as
yunyoo:tool-name, for exampleyunyoo:screenshot; - a command as
yunyoo subcommand, for exampleyunyoo observe screenshot.
If the text or a reference file mentions a tool, a subcommand or a host that the manifest does not declare, validation fails: a capability that is used without being declared bypasses the disclosure. All tools and subcommands are listed in the MCP tool reference and the Command-line reference.
Package limits
- The whole package is at most 2 MB, a single file at most 512 KB, with at most 200 files.
- The top level allows only the three files and three folders listed above, and the folders have no subfolders.
- File names use only lowercase letters, digits,
-,_and., and do not start with.. Text files are.md,.jsonor.txt, images must be PNG, and scripts must be.pyor.sh. - No symbolic links and no hard links.
- Text is UTF-8 without a BOM. The body of
SKILL.mdis at most 500 lines; the official skills keep it under 300 and put the finer detail inreferences/. - Packaging is reproducible byte for byte: packing the same source twice gives the same digest.
Safety model
A version appears in the market only after it passes these steps in order:
- Automatic scanning. Three independent scanners check the package in an environment without network access: our own rules, a leaked-secret check and a third-party skill scanner. If any one of them does not pass, the process stops.
- Evaluation. Both evaluations pass, see the next section.
- Human review. An administrator who did not take part in writing the skill reads every file and checks, item by item, the commands and the hosts it reaches, hidden or adversarial instructions, credential access and data leaving the device, and whether the declared permissions match what the skill actually did during evaluation.
- Signing. Only after the review passes is the package signature handed out with the version. The desktop app has the public key built in; before installing it verifies the signature and then the digest, and does not install if either does not match. The test environment and the production environment use different keys that do not work for each other.
When a problem is found after release, we withdraw that version: it disappears from the market and the public index, downloads are refused, desktop apps that have it installed show a notice, and for a serious problem the version is disabled.
A passed scan is not a safety endorsement. Scanners only find known patterns and evaluations only cover a limited set of tasks. They lower the risk; they do not guarantee that a skill is safe in every situation.
The following are not security boundaries, so do not rely on them:
- The text of a skill is not a permission. What the assistant can do is decided by your account permissions, the token scopes and the platform's checks. "Do not do this" in a skill is a request to the assistant, not a technical limit. The real limit is issuing a token with only the scopes and cloud phones it needs.
- What is on the screen is data, not instructions. The official skills tell the assistant not to follow instructions that appear on the screen, in web pages or in apps, but that depends on the assistant complying and cannot be treated as a defense.
- Scan and review results apply only to that version and that exact content. Other files with the same name, or an installed folder that was edited, are not covered; the desktop app marks installed folders that were changed.
How skills are evaluated
- Deterministic evaluation: without any AI, the command line runs every subcommand that appears in the skill's files and checks the exit code, the structure of the output and the time taken. This makes sure the commands in the text are real and work.
- Agent evaluation: an AI assistant completes a set of tasks with the skill, each task run 3 times. There are three kinds of tasks: ones where the skill should be used, ones where it should not be triggered, and boundary and compliance tasks (for example asking it to do something beyond its permissions, expecting it to refuse and say why). Using a tool it was told not to use, or acting when it should have refused, is not allowed even once.
- Both evaluations are compared with the platform's audit records: which tools were actually called and whether anything went beyond the declared scopes. A single call outside the scopes fails the evaluation.
