The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a runtime discovery mechanism to locate and execute internal components. Specifically, it instructions the agent to walk up the directory tree from the skill's location to find a specific source file (packages/baoyu-codex-imagegen/src/main.ts) and execute it using the bun runtime. This dynamic loading from computed paths is intended to support various local image generation backends but constitutes a dynamic execution pattern.
[COMMAND_EXECUTION]: The skill frequently utilizes shell commands and local scripts to automate the comic production workflow. This includes executing scripts/merge-to-pdf.ts to compile images into a PDF, and calling system utilities like sips and pngquant for image compression. The agent is also instructed to use shell probes such as test -f and test -d to manage project configurations and detect existing content.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection because its primary function is to ingest arbitrary user-provided text ('source content') to generate analysis documents and image prompts. The ingestion of this untrusted data is combined with significant capabilities, including local file writing and subprocess execution.
Ingestion points: The skill reads source content from local files or directly from user messages as described in workflow.md (Step 1.2).
Boundary markers: The instructions provide structure via markdown templates but do not implement specific delimiters or 'ignore' commands to isolate user content from the agent's logic during prompt generation.
Capability inventory: The skill performs extensive file writing (analysis.md, storyboard.md, multiple prompt files) and executes shell commands for document merging and image processing.
Sanitization: There is no explicit sanitization, validation, or filtering of the source content before it is processed by the LLM into actionable comic panels and prompts.

baoyu-comic