The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The file scripts/create-lora-config.md uses the ! syntax to execute shell commands at the time the skill is loaded. These commands include grep and find operations to extract project metadata and configuration details.
[CREDENTIALS_UNSAFE]: A dynamic context injection command in scripts/create-lora-config.md (grep -r "MODEL_PATH\|model_name" .env*) specifically accesses .env files. While intended to find model identifiers, .env files typically contain sensitive credentials and API keys.
[REMOTE_CODE_EXECUTION]: The rule rules/local-ollama-setup.md describes an installation process using curl -fsSL https://ollama.com/install.sh | sh. While the source is a well-known service (Ollama) and the documentation includes a warning to inspect the script first, this pattern involves executing remote code locally.
[DYNAMIC_EXECUTION]: Multiple scripts, including scripts/function-def.py and logic in rules/calling-validation.md, implement tool execution by parsing LLM-generated JSON strings and passing them dynamically to Python functions via func(**args). This is a standard but sensitive pattern for implementing function calling.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for processing external data for fine-tuning (rules/tuning-dataset-prep.md) and handling tool outputs (rules/calling-validation.md).
Ingestion points: Data is ingested from requirements.txt, pyproject.toml, .env files, and JSONL/JSON preference datasets.
Boundary markers: The instructions recommend using structured JSON output and strict schemas, which provide some delimitation.
Capability inventory: The skill has capabilities for file reading (Read, Glob, Grep), network requests (WebFetch), and shell execution (! commands).
Sanitization: While the skill recommends validation with a "judge model," there is no explicit code-level sanitization or escaping demonstrated for the raw data ingestion paths.

llm-integration