The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external tools via npx to perform specific tasks. This includes media-provenance for embedding metadata into images, replicate-mcp for local model serving, and a skill management command to install related Replicate skills. These commands are executed as part of the intended local development workflow.
[EXTERNAL_DOWNLOADS]: The script save-prediction.mjs downloads media artifacts (images, videos, audio) from URLs provided in Replicate's prediction records using the fetch API. Additionally, npx commands download necessary packages from the npm registry if they are not already present locally.
[SAFE]: The skill incorporates several security best practices to mitigate common risks associated with handling external data:
Secret Protection: It explicitly instructs the agent to check for the presence of REPLICATE_API_TOKEN without printing or logging its value, preventing accidental credential exposure.
Input Sanitization: The save-prediction.mjs script utilizes a slug() function to sanitize prompt strings and prediction IDs before using them in local filenames, protecting against path traversal or filename-based attacks.
Collision Prevention: The file-saving logic uses numeric versioning (__v002, etc.) and temporary files with random UUIDs to ensure that existing local files are never overwritten accidentally.
[SAFE]: The skill exposes an attack surface for indirect prompt injection as it processes data from the Replicate API to generate local artifacts. However, this is consistent with its primary purpose and includes sufficient sanitization.
Ingestion points: The save-prediction.mjs script processes prediction JSON objects (including prompts and output URLs) retrieved from the Replicate API.
Boundary markers: While explicit prompt delimiters are absent, the data is handled programmatically rather than being directly interpolated into a secondary prompt.
Capability inventory: The skill possesses file system access (writing and linking files), network access (downloading media), and the ability to spawn subprocesses via execFileSync.
Sanitization: Potentially malicious characters in prompts are stripped out by the slug() utility before being used in the file system.

replicate-local-workflow