The Agent Skills Directory

[PROMPT_INJECTION]: The skill overrides existing code safety constraints by instructing the agent to remove developer-placed warnings and suppressions. In references/comment-sicko.md, the subagent is told to 'kill' linter and TypeScript suppressions even if they 'protect correctness or safety'. It also instructs the agent to treat developer warnings such as IMPORTANT, do not remove, and too risky as 'scent' rather than valid exceptions, potentially causing the removal of critical safety-related documentation if the agent deems them not 'obvious'.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to malicious content within the files it processes.
Ingestion points: Processes caller-provided files and git diffs in SKILL.md (Step 1).
Boundary markers: Absent. The skill lacks delimiters to protect the agent from instructions embedded in code comments.
Capability inventory: The skill can delete code paths, modify parameters, and implement refactor fixes (Steps 3 and 4 in SKILL.md).
Sanitization: None. Maliciously crafted comments could be designed to look like 'confessions' or 'surprises' to trigger the persona's 'MUST KILL' logic, leading to the deletion of security-critical code or documentation.
[COMMAND_EXECUTION]: The skill orchestrates multiple sub-tools and agents (/how, /why, /architect) to inspect code logic and execute modifications, demonstrating significant autonomous control over the filesystem and potentially the tool execution environment.

no-comments