The Agent Skills Directory

[SAFE]: The skill consists entirely of markdown instructions and system prompts. There are no executable scripts (.py, .js, .sh), no remote package dependencies, and no shell commands included in the provided files. The instructions guide the host agent on how to use internal LLM capabilities rather than external tools.
[NO_CODE]: No code is shipped with this skill. It defines a workflow based on text-to-text transformation using specific prompt templates. This significantly reduces the attack surface for traditional code-based exploits.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted research paper content ({paper_content}). The provided prompts include explicit boundary markers ([PAPER CONTENT] and [END PAPER CONTENT]) and strict output formatting instructions ("Return only the markdown memo...") to mitigate the risk of content within the paper influencing the agent's behavior.
Ingestion points: The {paper_content} placeholder in dense-idea-prompt.md, experimental-log-prompt.md, and sparse-idea-prompt.md files.
Boundary markers: The prompts use bracketed delimiters to isolate external content.
Capability inventory: The skill is limited to generating markdown text files in a local bench/ directory. No network access or system command execution capabilities are defined or requested.
Sanitization: Relies on LLM instruction following and delimiter-based isolation.

paper-writing-bench