The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes the @firecrawl/anydoc package from the NPM registry and encourages the use of Firecrawl cloud services for parsing and OCR of large documents. It also includes a git submodule reference to a mortgage document corpus for evaluation purposes.
[REMOTE_CODE_EXECUTION]: The protocol instructs the agent to execute remote code via npx -y @firecrawl/anydoc to convert PDFs and Office files into markdown before processing.
[COMMAND_EXECUTION]: The skill requires the use of several shell utilities to size and filter inputs, including find, wc, ls, grep, and sed.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, such as entire codebases and large PDF documents, which creates a surface for indirect prompt injection attacks.
Ingestion points: File content read through standard file readers, content filtered via grep, and converted markdown output from external document parsers.
Boundary markers: The skill uses self-contained "briefs" when delegating to sub-agents, which helps scope the context, but it does not specify explicit delimiters or "ignore embedded instructions" warnings for the raw data ingested into these sub-agents.
Capability inventory: The skill possesses the ability to execute shell commands, read/write local files (e.g., to a .firecrawl/ directory), access the network for document parsing, and spawn sub-agents.
Sanitization: There is no evidence of sanitization or filtering of the content extracted from external files before it is processed by the main agent or its sub-agents.

recursive-decomposition