The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions include directives to override standard agent output behavior, specifically commanding the agent to omit reasoning, preambles, and Markdown fences when generating JSON output. It also contains instructions to disregard memory and avoid inferring answers from option positions, which are techniques used to override default model behavior.
[COMMAND_EXECUTION]: The skill directs the agent to perform various system-level tasks using shell commands and local scripts, including creating symbolic links with ln -snf, running preflight tests with make, and executing configuration management scripts such as scripts/link-config.sh.
[EXTERNAL_DOWNLOADS]: The skill references the use of bunx to add external agent skills (e.g., bunx skills add "$skill"). This utilizes a well-known package runner to fetch and install tools from external registries, which is a standard procedure for this repository's ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface as it ingests untrusted user-supplied "requested outcomes" and "evaluations" (SKILL.md) and possesses the capability to modify the file system and execute commands (references/change-routes.md).
Ingestion points: User-provided outcomes and multiple-choice evaluation questions.
Boundary markers: The instructions mandate reading specific reference files but do not establish clear boundaries or delimiters for untrusted user input.
Capability inventory: File system modification via symbolic links, execution of repository scripts and makefiles, and package installation via bunx.
Sanitization: There is no explicit evidence of input sanitization or validation to prevent malicious instructions within a "requested outcome" from manipulating shell commands.

dotfiles-change-router