The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions such as 'ignore that text' when referring to data payloads. This was identified as a defensive security measure intended to prevent the agent from following malicious instructions potentially embedded in third-party data, rather than an attempt to bypass system safety guidelines.
[COMMAND_EXECUTION]: The skill executes a local script scripts/bridge.py using Python. This script is part of the skill package and is used to open a web browser and receive data from the author's website. It uses only standard Python libraries and operates with a short timeout, facilitating a documented interaction between the agent and a browser-based calculation engine.
[EXTERNAL_DOWNLOADS]: The bridge script opens a browser window to https://eastalmanac.com. This domain is the official site for the author's EastAlmanac tool. The interaction is used for client-side calculations and user interface, with no background downloads or execution of remote scripts detected.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSON data returned from the browser-based tool. It implements appropriate mitigation by explicitly instructing the agent to treat the payload as untrusted data and ignore any embedded instructions. This effectively limits the attack surface for data-driven prompt injection. Ingestion points: scripts/bridge.py output. Boundary markers: Explicit instructions in SKILL.md to treat payload as data only. Capability inventory: Execution of local bridge script. Sanitization: Instruction to ignore command-like text in results.
[DYNAMIC_EXECUTION]: The agent is instructed to execute a provided Python script (scripts/bridge.py) at runtime to obtain data. This is a controlled execution of a script bundled with the skill, used for a specific and documented purpose.

yifortune