Agent skills
The repository ships four agent skills under skills/. They encode the parts of
this project that are easy to get wrong and expensive to discover — which SDK
call implements a feature, what must never be committed, what counts as a real
test note.
An agent that loads them contributes the same way a person following Adding camera coverage would. They are plain Markdown, so they are equally readable if you are doing the work yourself.
| Skill | Use it when |
|---|---|
alpha-sdk-doc-research |
Before implementing anything — research the SDK, camera help guides and Web API docs through the MCP servers |
alpha-sdk-add-api |
Implementing a missing endpoint, property or action mapping |
alpha-sdk-test |
Planning and running unit and end-to-end validation, including camera-backed curl checks |
alpha-sdk-publish |
Checking CI and public-repository readiness before pushing |
The order matters
Section titled “The order matters”Research first. The most common failure is implementing against a guess about what an SDK call does, then discovering the real semantics during hardware testing — by which point the endpoint shape is already wrong and the OpenAPI contract has been written around it.
alpha-sdk-doc-research exists to front-load that: it queries the indexed SDK
documentation and camera help guides so the design starts from what the SDK
actually provides, and from which bodies actually support it.
Documentation MCP servers
Section titled “Documentation MCP servers”The research skill relies on two MCP servers that index the SDK reference and
the camera help guides. Setting them up is covered in
MCP Server setup, and docs/MCP_SERVERS.md
carries the repository-side configuration.
Publishing safety
Section titled “Publishing safety”alpha-sdk-publish exists because this repository has an unusual constraint:
Sony’s SDK cannot be redistributed, so a routine commit can create a licensing
problem rather than a merge conflict.
The skill checks that no SDK source, binary, build folder, camera media or local credential is staged before a public push. CI enforces the same rule, but finding out locally is considerably better than finding out from a failed job on a public repository.