The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes task descriptions and notes from the Taskhold remote service and configuration from user-provided setup URLs. This content serves as instructions for the agent, creating a surface where malicious tasks could attempt to override agent behavior.
Ingestion points: The tools taskhold_list and taskhold_get in SKILL.md retrieve external data, and the instructions direct the agent to fetch a user-supplied setup URL.
Boundary markers: The instructions do not define specific delimiters or guardrails to isolate task content from the agent's core instructions, increasing the risk of the agent following instructions embedded within a task.
Capability inventory: The agent performs repository modifications and coding tasks based on the input and has the ability to write data back to the remote server via taskhold_add_note and taskhold_create.
Sanitization: No programmatic sanitization is present, although the skill provides natural language warnings to the agent not to store secrets in the task list.
[EXTERNAL_DOWNLOADS]: The skill fetches setup configuration and API keys from a URL provided by the user. It also communicates with taskhold.com to synchronize the task list. These network operations are intrinsic to the service provided by the vendor.
[DATA_EXFILTRATION]: The tools taskhold_add_note, taskhold_create, and taskhold_complete transmit local project context (such as task summaries and notes) to the remote Taskhold server. While this is the primary function of the skill, it establishes a data flow from the local environment to an external service. The skill includes explicit instructions to avoid putting secrets or large code blocks into tasks to mitigate this risk.

taskhold