The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute a local binary (referred to as a "checker") if the user provides an absolute path to the executable. While the instructions emphasize using parameter arrays rather than string concatenation or shell execution to prevent command injection, this capability allows the agent to run local files.
[DYNAMIC_EXECUTION]: The skill supports execution of binaries from paths provided dynamically by the user at runtime. The instructions in references/language-and-fact-check.md specify a fixed CLI interface (<checker> check --document <path> ...) to manage this process, which represents dynamic execution from computed paths.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant ingestion surface, as it is designed to read and process existing project documentation and user-provided technical content.
Ingestion points: SKILL.md (Stage 1) and references/knowledge-base-organization.md instruct the agent to read AGENTS.md, CLAUDE.md, file indices, and other project-related documents.
Boundary markers: references/language-and-fact-check.md defines "Technical String Protection" rules to safeguard fenced code, paths, and commands from being modified by the language processing phase.
Capability inventory: The skill can read project files, write new Markdown documentation, and execute local binaries via subprocess calls (in the validation scripts).
Sanitization: It implements a "Fact Gate" to identify and protect sensitive literals and evidence before prose optimization occurs.
[PRIVILEGE_ESCALATION]: The skill instructions and test cases (e.g., in evals/evals.json) involve the creation of runbooks and guides that may contain privileged commands such as sudo systemctl restart. While the skill includes instructions to preserve these commands exactly and not to add unauthorized ones, the agent acts as a conduit for documenting and potentially facilitating the execution of elevated commands.

custom-skills-doc-writer