The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a prompt template in SKILL.md under the 'Tool-Testing Agent Pattern' section intended to optimize tool descriptions based on execution history.
Ingestion points: The template interpolates {failure_examples}, which typically consists of logs from previous agent interactions that may contain untrusted external data.
Boundary markers: The template lacks explicit delimiters or instructions to ignore embedded commands within the failure examples.
Capability inventory: While the skill itself is instructional, the patterns it describes (such as the bash sandbox) represent significant capabilities if implemented.
Sanitization: No sanitization or escaping logic is suggested for the interpolated data.
[COMMAND_EXECUTION]: The skill documentation in SKILL.md and references/architectural_reduction.md promotes the 'Architectural Reduction' pattern. This pattern encourages providing agents with low-level command-line access (e.g., execute_command for running bash) rather than restricted specialized tools. While the skill correctly identifies the need for sandboxing, this design philosophy advocates for granting agents high-privilege primitives that increase the overall attack surface.

tool-design