The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted external sources, specifically CAD drawings in DXF, DWG, and SVG formats.
Ingestion points: Untrusted data enters the agent context through file-reading tools such as cadcraft-cli info, inspect_drawing, find_text, and query_entities (SKILL.md).
Boundary markers: The skill lacks explicit instructions for using delimiters or boundary markers to segregate file content from instructions, though it notes the texts tool removes formatting.
Capability inventory: The agent has extensive capabilities including file system writes (save, export, run --save), execution of 427 engine commands, and interactive control of the GUI via the control channel (SKILL.md).
Sanitization: While the skill mentions the craft-docpath library for lexical path validation to prevent unauthorized network access, there is no evidence of semantic sanitization for text content extracted from entities.
[COMMAND_EXECUTION]: The skill relies on the execution of the cadcraft-cli binary and engine-level commands to perform its core functions. This includes headless processing, format conversion, and bridging to a running application instance.
[DYNAMIC_EXECUTION]: The skill supports the generation and execution of AutoLISP (.lsp) and script files (.scr). This involves assembling and loading executable logic at runtime to automate drafting tasks.
[DATA_EXFILTRATION]: The skill implements a JSON control channel communicating over 127.0.0.1 (localhost). While this is a network operation, it is limited to local inter-process communication for application control.

cadcraft