The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from repository markdown files and CI failure logs to identify broken links and generate repair commands. An attacker could inject malicious instructions into these sources to manipulate the audit outcome or influence the generated repair logic.
Ingestion points: CI failure logs (baseline.log) and markdown files in the repository.
Boundary markers: No specific delimiters or safety instructions are used when interpolating external log data into agent prompts.
Capability inventory: Extensive use of shell commands including sed, git, gh, and curl for file manipulation and network interaction.
Sanitization: The methodology relies on a manual 'Verification-First Rule' and human review checkpoints, but does not provide programmatic sanitization of ingested content.
[COMMAND_EXECUTION]: The skill instructions include a task to locate and modify its own files on the local system, which is a form of self-modification.
Evidence: The post-merge cleanup playbook (post-merge.md) includes an agent task (agent-D) that attempts to find the skill's installation directory (e.g., ~/.claude/skills/) and append rules to the skill's markdown files.
[EXTERNAL_DOWNLOADS]: The skill uses tools that fetch external content during execution.
Evidence: Recommends using npx markdown-link-check which downloads packages from the npm registry if not present. It also uses curl to probe arbitrary external URLs found during the audit process.

alterlab-link-health