The Agent Skills Directory

[SAFE]: All Python code samples provided in the skill utilize standard libraries such as dataclasses, string, and json. The implementation of the PromptTemplate and ConstructionPromptLibrary classes is transparent and lacks any hidden execution logic or dangerous subprocess calls.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external user data by interpolating it into prompts for an LLM, creating a potential surface for indirect prompt injection.
Ingestion points: Data enters the skill context through input_variables such as line_items, schedule_data, and element_data as defined in SKILL.md.
Boundary markers: The templates use plain-text headers (e.g., 'Line Items to Classify:') to separate user data from instructions but do not employ explicit safety delimiters or 'ignore' instructions for the interpolated content.
Capability inventory: The skill manifest in claw.json requests filesystem access, which is necessary for its stated purpose of reading construction documents (CSV, Excel, JSON).
Sanitization: The PromptTemplate.format method uses string.Template.safe_substitute, which performs variable replacement without sanitizing or escaping the content for potentially malicious LLM instructions.

prompt-templates