The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from repository files and code diffs, creating a potential vector for instructions embedded within the source code to influence agent behavior.
Ingestion points: The critic prompt in references/comment-reviewer.md specifically requests "parent scoped files or diff" or a "current diff" as its primary input for analysis.
Boundary markers: The instructions do not define clear delimiters or "ignore previous instructions" safety markers for the content being reviewed, which could lead the model to follow natural language directives found inside comments.
Capability inventory: The skill is capable of modifying the local filesystem by applying comment deletions and identifying refactor targets as directed by the critic findings.
Sanitization: There is no evidence of sanitization or structural validation performed on the ingested code or diff content.
[PROMPT_INJECTION]: The critic agent defined in references/comment-reviewer.md adopts an aggressive persona ("Comment Sicko") and uses absolute, emotive directives such as "I hate comments", "Everything else is meat", and "That list is my only leash". This type of role-play, while functionally scoped to comment cleanup, mimics techniques used to override standard model constraints.

pstack-no-comments