The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the bun runtime to execute internal scripts (scripts/generate.mjs, scripts/validate-portable-v1-output.mjs) for artifact rendering and validation. These operations are limited to the local workspace.
[DYNAMIC_EXECUTION]: The skill generates Python code for plugin artifacts. This capability is guarded by a strict manifest schema and regex-based input filtering that prevents the injection of malicious logic or sensitive data into the generated files.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a surface for generating instructions and code from user prompts. It mitigates injection risks through a multi-layered defense:
Ingestion points: Natural language requests for artifacts are processed in SKILL.md and normalized into a JSON manifest.
Boundary markers: Explicit routing rules in rules/routing.md and rules/write-safety.md define strict scope limits, while the instruction set mandates the rejection of out-of-scope tasks.
Capability inventory: The skill possesses local file writing and directory management capabilities through its bundled scripts.
Sanitization: All user-provided strings are sanitized using the safeSummary regex in scripts/generate.mjs and validated against assets/manifest.schema.json, which block special characters and sensitive keywords like 'secret', 'token', and 'discord'.
[SAFE]: The skill implements robust write-safety protocols, including ownership markers (.hermes-agent-maker/ownership.json) and transactional journals, ensuring that it only modifies files it created and can recover from interrupted operations without leaving the workspace in an inconsistent state.

hermes-agent-maker