The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is designed to fetch content from multiple external sources. It retrieves a skill registry from https://raw.githubusercontent.com/heilcheng/awesome-agent-skills/main/README.md and uses git clone or the GitHub API to download repository contents from various individual users, such as coffeefuelbump, chrisvoncsefalvay, lackeyjb, zxkane, and smerchek.
[REMOTE_CODE_EXECUTION]: The instructions mandate that the agent "Test any relative imports" and "Always test skills before committing" for the newly imported content. This workflow could lead to the execution of unvetted scripts or code contained within the third-party repositories during the validation phase.
[COMMAND_EXECUTION]: The skill frequently uses shell commands such as git clone, git sparse-checkout, git submodule, and curl to manipulate the local file system and synchronize with remote repositories.
[DATA_EXFILTRATION]: Because the skill possesses both network access (via WebFetch and curl) and the ability to read from the repository, a malicious imported skill being executed during the "test" phase could potentially exfiltrate sensitive configuration files or repository data.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface. It ingests untrusted data from the awesome-agent-skills registry and individual SKILL.md files.
Ingestion points: SKILL.md (Workflow steps for fetching and parsing).
Boundary markers: Absent. No specific delimiters or instructions to ignore embedded commands are present.
Capability inventory: Extensive, including git push, file writes, and the general "test/validate" instruction across SKILL.md.
Sanitization: Absent. The agent is instructed to normalize metadata but not to sanitize the instructions or scripts within the imported skills.

skills-manager