The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple external binaries and shell scripts located on the user's file system, specifically rmbc-update-check, rmbc-analytics, and rmbc-workspace. These paths are dynamically resolved based on the presence of the dtc-copywriting-skills directory in the user's home or local environment.
[COMMAND_EXECUTION]: The preamble block uses sed -i to modify the user's ~/.rmbc-skills/config.yaml file to enable or disable analytics based on the user's response to an AskUserQuestion prompt.
[EXTERNAL_DOWNLOADS]: The skill triggers network operations indirectly by calling rmbc-update-check and rmbc-analytics. Additionally, it uses the open command to launch a browser to a YouTube URL, which, while benign in this context, demonstrates a capability to trigger external actions.
[DATA_EXFILTRATION]: The skill collects and transmits usage data (skill name, product identifiers, and usage tier) to a remote service via the rmbc-analytics tool. Although this is gated by a telemetry opt-in, it represents an automated data transmission mechanism.
[PROMPT_INJECTION]: The skill ingests untrusted user inputs (such as initial_product, upsell_product, and target_audience) and interpolates them directly into prompts without boundary markers or sanitization. This creates an attack surface for indirect prompt injection if the user-supplied data contains malicious instructions.
[DATA_EXFILTRATION]: The skill reads internal configuration and state files (config.yaml, .intro-seen, .telemetry-prompted) and project-specific research files (research.md, mechanism.md, brief.md) to influence its behavior and logging.

upsell-sequence-writer