Use when
Use this agent when a user wants to turn one readable image containing exactly one visually dominant person or pet into a cursor-driven native macOS or Windows desktop-pet application. The user must also provide an explicit desktop-pet style.
Do not use this agent for multi-subject source images, conventional websites, Electron wrappers, Linux desktop packages, or unrelated media editing.
Prompt context requirement
Every prompt sent to this agent must include the destination computer system as one normalized line:
```text client_os: macos ```
or:
```text client_os: windows ```
`client_os` describes the user's local computer that will run the final package. It never describes the Agent, OpenClaw, sandbox, browser, or build host. Do not omit this line, infer it from the Linux sandbox, or substitute values such as `darwin`, `win32`, or `linux`. Normalize an explicitly known Mac/Windows value before sending the prompt. If the user's computer system is genuinely unknown, obtain it before starting generation.
- One accessible image containing one visually dominant core subject; one person or one pet is supported.
- One explicit, non-empty desktop-pet style.
- `client_os: macos` or `client_os: windows`.
The agent does not invent a style or guess the destination system. It asks only for missing required input and creates no run artifacts until all three inputs are present.
Outputs
- For `client_os: macos`: a ZIP containing a native universal `arm64` + `x86_64` AppKit `.app`.
- For `client_os: windows`: an unsigned native Windows 10/11 x86-64 NSIS Setup EXE.
Both applications display the generated character as a transparent always-on-top overlay, map horizontal pointer position to the complete main animation in both directions, loop an intro before interaction, and play a one-shot outro after reaching the end and settling. Generated packages have no runtime network dependency.
Processing contract
- Produce exactly five coherent pose images and four adjacent image-to-video transitions.
- Preserve the single subject's identity, subject type, anatomy, clothing or coat, markings, proportions, and distinctive accessories.
- Generate `1920x1080`, `16:9` pose images against one deterministic chroma key for the whole run: green `#00FF00`, blue `#0000FF` when the subject contains significant green, or magenta `#FF00FF` when it contains significant green and blue. Gray backgrounds are unsupported.
- Generate four Seedance 2.0 transitions at `480p`, approximately five seconds each.
- Create exactly three transparent RGBA frame packs: `intro`, `main`, and `outro`.
- Select one-pixel matte erosion first and use two pixels for the complete run only when sampled residual-outline measurements justify it. Never vary erosion strength frame by frame.
- Reuse the same validated transparent frame pack for both operating-system branches.
- Assemble only the matching immutable precompiled runtime; do not compile Swift or C++ during an ordinary user run.
- Validate metadata, runtime compatibility, frame names and counts, PNG alpha, licenses, capability records, and package integrity before delivery.
### macOS
Use the versioned universal AppKit Runtime v2, apply an ad-hoc signature with `rcodesign`, and deliver a ZIP. It is not Developer ID notarized, so first launch may require right-clicking the app and choosing **Open**.
### Windows
Use the versioned native Win32 x86-64 Runtime v1 and rebuild the NSIS installer around the generated AppConfig and complete frame pack. Do not patch an already-built installer. The package is unsigned, so Windows SmartScreen may show a warning.
Limitations
- Images that are missing, corrupt, or contain multiple similarly prominent candidate subjects cannot proceed.
- A clear, well-lit, full-body or near-full-body image with minimal obstruction produces the best result.
- External image and video generation may introduce identity, anatomy, framing, or motion defects; retries remain bounded by the skill contract.
- Linux is the assembly environment, not a supported destination desktop runtime.
- Per-run macOS or Windows GUI launch is not claimed from the Linux sandbox; deterministic structural validation is mandatory.
- Credentials, source videos, build tools, runtime source, caches, and unnecessary intermediate artifacts are excluded from final packages.
