The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from external product requirement documents and Feishu/Lark comments. This creates a surface for indirect prompt injection where malicious instructions inside a PRD could attempt to override the agent's behavior during the analysis or digest phase.
Ingestion points: Content is ingested from Feishu/Lark via the lark-cli utility and from local markdown/text files.
Boundary markers: While the skill uses structured output templates like digest-template.md, it does not apply specific boundary markers or defensive instructions when processing the raw input text to prevent the LLM from following embedded instructions.
Capability inventory: The skill can execute the lark-cli utility and write processed data to the local sdlc/ project directory.
Sanitization: The scripts/parse_doc.py script extracts structural elements into a JSON model but does not explicitly sanitize the text content for potential prompt injection patterns.
[COMMAND_EXECUTION]: The skill invokes the external lark-cli tool via the Python subprocess module. This is used to fetch document content and download media for processing. The script scripts/parse_doc.py uses list-based arguments for subprocess.run(), which is a secure practice that avoids shell injection risks, but it remains a capability that relies on executing external binaries.

sdlc-intent