The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where untrusted inputs, such as user prompts and existing local source code, are processed by a third-party model whose output is then executed by the agent.
Ingestion points: The helper script scripts/call-deepseek.js reads file content using fs.readFileSync; SKILL.md instructs the host to aggregate project context using file-reading tools.
Boundary markers: Prompt templates in SKILL.md include instructions for DeepSeek to format output inside markdown code blocks.
Capability inventory: The skill workflow utilizes write_to_file, replace_file_content, and run_command to apply model-generated changes.
Sanitization: The skill relies on the agent's instructions to verify the generated code's completeness before execution, but does not implement automated sanitization.
[COMMAND_EXECUTION]: The skill provides a helper script and instructional steps for executing shell commands and tool actions to bridge text generation with filesystem changes.
Evidence: SKILL.md describes using run_command for testing and validation of code generated by DeepSeek.
Evidence: scripts/call-deepseek.js is a utility script that queries a local API gateway and outputs results to stdout, designed for use within automated command pipelines.

delegate-to-deepseek