The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates task scheduling by creating and loading macOS LaunchAgents (~/Library/LaunchAgents/ai.<slug>.loop.exec.plist) using the launchctl utility. This allows loops to maintain persistence and run unattended.
[EXTERNAL_DOWNLOADS]: The instructions direct the user or agent to install a global CLI tool, looptop, from the public NPM registry (npm install -g looptop). This external tool is central to the skill's monitoring and control capabilities.
[DATA_EXFILTRATION]: The workflow requires synchronizing state, tickets, and findings with external services including Linear and GitHub. Additionally, the skill scans local sensitive paths, such as the Obsidian vault registry (~/Library/Application Support/obsidian/obsidian.json), to identify project state storage.
[PROMPT_INJECTION]: The documentation explicitly addresses "Injection propagation," a vulnerability where untrusted data from ticketing systems or web applications can poison the agent's context.
Ingestion points: Linear tickets, GitHub issues, and application data discovered during "free-roam" testing phases.
Boundary markers: The skill recommends using "frozen specs" and "one-task-per-iteration" scoping to mitigate influence.
Capability inventory: The system has access to CLI tools (npm, launchctl, gh), file system writes, and external network APIs.
Sanitization: It advises inspecting content at communication boundaries and treating tool output as untrusted data.

software-factory