The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, specifically pull request bodies, linked tickets, and git diffs, which are potential vectors for indirect prompt injection.
Ingestion points: Data is gathered from pull request titles, bodies, linked tickets, and review outcomes as part of the evidence gathering process defined in SKILL.md.
Boundary markers: The procedure explicitly instructs the agent to read the diff before the pull request description to avoid being anchored to the author's potentially biased framing. While specific XML-style delimiters are not used, the logic serves as a functional boundary.
Capability inventory: The skill uses Bash, Read, Grep, and Glob to perform repository-level operations such as git show and git diff.
Sanitization: The output is strictly defined by a JSON contract (references/output-contract.md), which ensures that descriptions are structured and behavioural claims are anchored to specific line ranges and commit SHAs, reducing the risk of malicious instructions influencing the agent's output format.
[SAFE]: The skill identifies external companion tools available via npx skills add crissmoldovan/agent-skills. This is a user-facing installation instruction referencing the skill's own author and does not involve the automatic execution of remote code at runtime.

describe-changes