The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill advocates for an 'Architectural Reduction' pattern (documented in SKILL.md and references/architectural_reduction.md) that replaces specialized tools with powerful primitives like execute_command. This increases the ingestion surface for untrusted data processed via system utilities.
Ingestion points: The skill suggests tools like grep, cat, and find should be used directly on the filesystem (as noted in references/architectural_reduction.md).
Boundary markers: No specific boundary markers or instruction-ignoring delimiters are mentioned for the output of these primitive tools.
Capability inventory: The design pattern encourages providing broad filesystem access (ls, find, cat, grep) and direct database execution (execute_sql).
Sanitization: The instructions do not emphasize sanitizing or escaping the output of these commands before the agent processes them, which could lead to an agent following instructions embedded within a read file or database result.
[COMMAND_EXECUTION]: The skill explicitly recommends patterns where the agent is given a tool to execute system-level commands (execute_command) to browse the filesystem (e.g., grep, find, ls). While described as a 'sandbox' in the case study, this pattern inherently grants higher autonomy and broader access to the underlying environment compared to restricted APIs.

tool-design