The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external documents (technical articles, READMEs, blog posts) and delegates them to subagents for processing. This creates a surface for indirect prompt injection if those documents contain adversarial instructions intended to influence the subagent's behavior.
Ingestion points: Processes Markdown input from stdin and references external source files (source data, research results, reference documents) via their paths.
Boundary markers: The skill structures requests into sections like Objective, Scope, and Constraints, though the effectiveness depends on the subagent's adherence to these boundaries.
Capability inventory: The skill can execute shell scripts, read local files, and communicate with external AI providers (Anthropic, OpenAI, Devin, OpenCode).
Sanitization: Not explicitly present in the shell wrappers; logic is deferred to the bundled Node.js CLI tool.
[DYNAMIC_EXECUTION]: All primary interface scripts (e.g., run.sh, prepare.sh, dispatch.sh) act as shims that dynamically execute a bundled Node.js script (delegate-cli.mjs) using a path computed relative to the script's directory.
[EXTERNAL_DOWNLOADS]: The skill references several external resources and services for model pricing and execution.
Evidence: model-token-prices.json contains links to pricing pages for Anthropic, OpenAI, Cognition (Devin), Cursor, and OpenCode. These are well-known technology and AI services.

delegate-prose