The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external documents (PDFs, images) through the LandingAI ADE service. This creates a surface for indirect prompt injection if a malicious document contains instructions designed to manipulate the extraction or classification LLM.
Ingestion points: Files are ingested from local paths provided to the agent (e.g., doc_path in SKILL.md, input_dir in batch-processing.md).
Boundary markers: The skill uses regex (re.sub(r"<a id='[^']*'>\s*</a>", "", text)) to strip ADE-specific anchor tags from text before processing, but lacks comprehensive delimiters for the extracted content itself.
Capability inventory: The skill has capabilities for network operations (LandingAI, OpenAI, Snowflake APIs), file system writes (Path.write_text, img.save), and subprocess execution via pytesseract to invoke the tesseract binary.
Sanitization: Use of structured Pydantic schemas (e.g., InvoiceSchema in schema-catalog.md) provides significant mitigation by forcing the model output into a strictly defined format, reducing the efficacy of many injection attacks.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands and Python snippets for diagnostic purposes (e.g., Tool 1 and Tool 2 in SKILL.md). These snippets involve file system access and rendering operations.
[DYNAMIC_EXECUTION]: The diagnostic tools utilize the `python
<< 'EOF'` pattern to execute multi-line Python scripts dynamically from the shell. While used for legitimate exploration, this is a form of dynamic execution that relies on the agent's ability to safely construct and run code blocks.

document-workflows