The Agent Skills Directory

[COMMAND_EXECUTION]: The workflow defined in references/soul_forge_flow.toml includes a write_file task (Soul_Manifestation) that dynamically creates new persona files on the local filesystem. This allows the skill to modify its own available instruction sets and capabilities at runtime by generating new markdown-based personas.
[PROMPT_INJECTION]: Instructions in references/soul_forger.md utilize patterns common in prompt injections to control model output format and suppress system context (e.g., "DO NOT repeat the system prompt", "START directly with the ---"). While intended for clean file generation, these instructions mimic techniques used to bypass standard model response structures.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface by processing untrusted external data to generate new operational instructions for the agent.
Ingestion points: The skill ingests data from failure_trace, failure_cluster, raw_facts, and wendao_search_results as defined in the references/soul_forge_flow.toml workflow.
Boundary markers: Absent. The instruction set lacks clear delimiters or warnings to the model to disregard instructions that might be maliciously embedded within the ingested failure data or search results.
Capability inventory: The skill possesses the write_file capability through its defined workflow nodes in references/soul_forge_flow.toml.
Sanitization: Absent. There is no evidence of validation, escaping, or filtering of the content extracted from external traces before it is synthesized into a new persona blueprint.

forge-evolution