The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates the installation of agent skills by executing a bundled script (scripts/sync-symlinks.sh) to manage two-tier symlink chains in the user's home directory. Additionally, the 'diagnose' mode generates unified diffs and applies them to the source code of other skills via the git apply command.
[DYNAMIC_EXECUTION]: In its diagnostic workflow, the skill creates and applies code modifications to local skill files. This dynamic generation and application of logic is the primary function of the 'diagnose' mode; it is mitigated by a mandatory confidence-scoring step (requiring a score of at least 90%) and requires explicit user confirmation before any filesystem changes are performed.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection because its core workflows involve reading and analyzing external, potentially untrusted data, including other skills' source code, session transcripts, and execution logs. If these inputs contain malicious instructions intended to mislead the LLM, they could influence the generated audit reports or diagnostic patches. The skill includes internal documentation advising on the use of boundary markers and sanitization to reduce this risk.

create-skill