The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements workflows that interpolate external data (such as questions and model responses) directly into prompts for self-critique and revision processes.
Ingestion points: Untrusted data enters the agent context via the prompts list and generated model outputs in the supervised learning and RLAIF workflows (SKILL.md).
Boundary markers: The templates provided (e.g., critique_prompt, revision_prompt, preference_prompt) use simple string formatting (e.g., {question}, {response}) without using delimiters or instructions to the model to ignore any instructions embedded within the interpolated data.
Capability inventory: The skill employs powerful capabilities, including text generation via Hugging Face transformers and model fine-tuning using the trl library.
Sanitization: There is no evidence of input validation, escaping, or filtering of the content provided in the variables before they are processed by the LLM.

constitutional-ai