The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions create a surface for indirect prompt injection by encouraging the agent to ingest and analyze external, untrusted data artifacts.
Ingestion points: SKILL.md instructs the agent to read "captured traces," "network requests," "event logs," "HAR files," and "log dumps" to help diagnose bugs (Phase 1, items 5 and 10).
Boundary markers: There are no instructions to use specific delimiters or boundary markers when the agent processes this external content to distinguish data from instructions.
Capability inventory: The skill allows for significant local capabilities including executing shell commands (bash), performing network operations (curl), and running arbitrary scripts (python, node, Playwright).
Sanitization: While the skill provides instructions to redact secrets like API keys, it lacks safeguards against malicious textual instructions embedded within the logs or traces it processes.
[COMMAND_EXECUTION]: The skill's primary function is to have the agent write and execute various scripts and CLI commands (e.g., curl, bash, git bisect) to create deterministic reproduction loops. This behavior is expected for a debugging skill but requires the environment to have appropriate command execution permissions.
[DYNAMIC_EXECUTION]: The skill requires the agent to dynamically generate code for "Throwaway harnesses," "test loops," and "instrumentation" in multiple languages (Phase 1 and Phase 4). This runtime code generation and execution is the core mechanism of the skill.

diagnosing-bugs