The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to manage the visibility of technical metadata, such as directing the agent to write YAML configuration blocks to a manifest file instead of displaying them in the chat interface. This is a design choice for user experience rather than an attempt to bypass safety controls or hide malicious intent.
[COMMAND_EXECUTION]: The skill utilizes command-line operations and Python modules, such as python -m examples.sft.muse_glimmer_starter, to facilitate machine learning workflows. These actions are tied to official vendor cookbook examples and represent the core intended functionality of the assistant.
[DATA_EXFILTRATION]: Explicit privacy safeguards are included, instructing the agent to never send raw customer prose, datasets, credentials, or paths in telemetry. It also provides secure guidance for users to manage their API keys using environment files rather than inputting them directly into the chat context.
[EXTERNAL_DOWNLOADS]: The skill references and utilizes the official Fireworks AI cookbook repository for code and task recipes. These are recognized vendor resources and are used to provide the user with pre-vetted machine learning starters.

research