The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data retrieved from remote repositories and local documentation.
Ingestion points: Reads live files (README, ARCHITECTURE.md, etc.) and metadata from the rlaope/oh-my-hermes GitHub repository.
Boundary markers: Requires distinct labeling of public vs. local facts and citation of specific commit/version IDs.
Capability inventory: Performs file reads in the local OMH home directory and shell command execution using the omh CLI.
Sanitization: Includes a specific instruction to ignore any commands or prompts embedded within the retrieved documentation to prevent unintended execution.
[COMMAND_EXECUTION]: Executes several sub-commands of the omh CLI to inspect the local installation state.
Evidence: Utilizes omh list, omh status, omh doctor, and omh model-chains.
Control: Explicitly identifies the side effects of omh doctor and provides specific flags like --no-state-write and --no-personal-files to limit the command's access to user data.
[EXTERNAL_DOWNLOADS]: Pulls product documentation from an external repository to ensure information is current.
Evidence: Connects to github.com/rlaope/oh-my-hermes to retrieve repository descriptions, branch information, and official source files.

omh-docs