The Agent Skills Directory

[PROMPT_INJECTION]: The skill generates automation scripts (ralph.sh, ralph-once.sh) that ingest data from an external, potentially untrusted source (GitHub issues).
Ingestion points: The generated scripts are instructed to "Look at the open GitHub issues" and pick tasks based on issue descriptions and comments (SKILL.md, Step 4 and 5).
Boundary markers: The prompt templates provided for the generated scripts lack explicit boundary markers or instructions for the agent to ignore embedded instructions within the ingested issue text.
Capability inventory: The scripts invoke the claude CLI with instructions to use the /execute tool, which typically possesses file system access and subprocess execution capabilities.
Sanitization: No sanitization or validation logic is defined to filter malicious instructions that might be embedded in GitHub issue content processed by the automated loop.
[COMMAND_EXECUTION]: The skill generates and facilitates the execution of local shell scripts to manage automation loops.
The skill creates and makes executable ralph.sh and ralph-once.sh scripts (SKILL.md, Step 7).
It optionally modifies package.json to include entry points for these scripts, facilitating their invocation within the project environment.
While these actions are intended for infrastructure setup, they establish a mechanism for running dynamically generated logic that acts on external inputs.

setup-ralph-loop