The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The script file scripts/create-lora-config.md utilizes dynamic context injection syntax (!command) to execute shell commands when the file is loaded by the agent. Specifically, it performs a recursive grep across .env files to discover model configurations. Since .env files are standard locations for sensitive credentials, API keys, and environment secrets, automated reading of these files poses a significant risk of data exposure.
[DYNAMIC_EXECUTION]: The training script scripts/dpo-training.py explicitly enables trust_remote_code=True when loading models and tokenizers via the Hugging Face Transformers library. This configuration permits the execution of arbitrary Python code defined within the model repository being accessed, potentially leading to unauthorized code execution if an untrusted or compromised model is loaded.
[INDIRECT_PROMPT_INJECTION]: The skill implements a template system in scripts/create-lora-config.md that interpolates user-provided input ($ARGUMENTS) directly into markdown and YAML structures. The lack of explicit boundary markers or sanitization for this external data creates an attack surface for indirect prompt injection if the resulting content is subsequently processed by an LLM.
[EXTERNAL_DOWNLOADS]: The documentation in rules/local-ollama-setup.md provides instructions to download and execute an installation script from ollama.com. This represents a remote code execution pattern, although it targets a well-known service and includes a recommendation for the user to inspect the script before execution.

llm-integration