The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection. It instructs the agent in SKILL.md to load and apply instructions from the ~/.claude/PAI/USER/SKILLCUSTOMIZATIONS/Science/ directory to override default behavior.
Ingestion points: Files in ~/.claude/PAI/USER/SKILLCUSTOMIZATIONS/Science/ (e.g., PREFERENCES.md).
Boundary markers: None provided; the agent is simply told to apply the found resources.
Capability inventory: The skill workflows include execution of shell commands via curl, bun, docker, and psql (seen in SKILL.md, Workflows/DesignExperiment.md, and Examples.md).
Sanitization: No validation or sanitization is performed on the content of the loaded files.
[COMMAND_EXECUTION]: SKILL.md requires the execution of a curl command in the background (&) targeting localhost:8888 immediately upon skill invocation. This command is configured to suppress output (> /dev/null 2>&1).
[COMMAND_EXECUTION]: Workflows/DesignExperiment.md instructs the agent to run a local script using the bun run command.
[COMMAND_EXECUTION]: Examples.md provides templates for executing infrastructure and database commands like docker and psql.

Science