The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill includes embedded Python scripts (in SKILL.md and references/system-prompt-rebuild.md) used to modify config.yaml and SOUL.md files at runtime. These scripts perform complex file I/O and YAML manipulation to update agent system prompts.
[COMMAND_EXECUTION]: Extensive use of shell commands and the hermes CLI is documented to automate the management of agent profiles and their configurations.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a significant attack surface by facilitating the modification of agent.system_prompt across all swarm profiles.
Ingestion points: The new_string parameter in patch calls and the new_prompt_with_plain_newlines variable in Python scripts.
Boundary markers: Absent; there are no delimiters or warnings to ignore embedded instructions in the updated content.
Capability inventory: The skill uses patch, read_file, hermes config set, and direct Python file writes to achieve its goals.
Sanitization: Absent; the skill performs direct string replacement and YAML overwriting without validation or filtering of the content being injected into the system prompts.
[PRIVILEGE_ESCALATION]: The instructions explicitly describe methods to bypass "cross-profile soft guards" intended to restrict writing to other agents' directories. This includes using a cross_profile=true flag and a Python-based shell bypass that directly writes to files, circumventing platform-level tool restrictions.

flutter-soul-stewardship