The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes repository source files, dependency manifests, and existing automation scripts to generate CI/CD workflows, which creates a surface for indirect prompt injection if an attacker embeds malicious instructions in the codebase.
Ingestion points: references/generate.md (Phase 1) performs a full analysis of the repository's source code and manifests.
Boundary markers: The skill requires an interactive "analyze-then-confirm" conversation, forcing the agent to present findings and policy choices to the user before generating code.
Capability inventory: The skill is designed to create and commit GitHub Actions workflows and modify repository settings via CLI tools.
Sanitization: references/security.md explicitly instructs the agent to avoid shell interpolation of untrusted data (like PR titles) by passing them through environment variables.
[COMMAND_EXECUTION]: The skill utilizes system and platform-specific CLI tools to gather environment data and apply hardening settings.
Evidence: Uses git remote get-url to detect the hosting platform and gh api to read or update repository security policies.
Safety Context: The instructions include a specific sed filter to strip potential credentials from the git remote output before it is processed or echoed, preventing accidental secret exposure.
[EXTERNAL_DOWNLOADS]: The skill references external resources and well-known tools for its functionality and recommendations.
Evidence: Inherits prompt logic from the vfarcic/dot-ai repository and recommends standard security and optimization tools such as zizmor, actionlint, and Renovate.
Vendor Resources: Directs users to the author's own repository (vtmocanu/skills) for feature requests and issue tracking.

cicd-expert