The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted text (such as tool descriptions and system prompts), which creates a surface for indirect prompt injection.
Ingestion points: The skill ingests text via the omp compress <file> command or through direct text input described in the compression procedure (SKILL.md).
Boundary markers: The documentation explicitly mentions that source content is "quoted inside a nonce-delimited block and declared inert" to prevent the model from obeying instructions embedded within the text being compressed.
Capability inventory: The skill utilizes tools like rewrite and approve and has the capability to write output to the file system (using -o or -i flags).
Sanitization: The verification process requires the agent to declare every "loss" (deleted content), which provides a mechanism for human review of any skipped or potentially malicious instructions.

semantic-compression