The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external text inputs including user-defined narratives, research topics, and style reference sources (which can be remote HTTP/HTTPS URLs or arXiv IDs) to generate paper plans and LaTeX sections via LLM calls (gpt-6-astra through Codex MCP). Adversarial content embedded in these inputs could manipulate the pipeline's behavior or outputs.
Ingestion points: Ingests user-supplied $ARGUMENTS, NARRATIVE_REPORT.md, and external reference files passed to the --style-ref option.
Boundary markers: Absent; the system does not wrap or isolate ingested data with strict semantic boundaries or specific instruction-guarding prompts.
Capability inventory: Involves local file modifications, execution of helper scripts (extract_paper_style.py, verify_paper_audits.sh), document compilation (latexmk), and interacting with Codex MCP servers.
Sanitization: No input validation or sanitization is specified for raw narrative content or text retrieved from external style sources.
[COMMAND_EXECUTION]: The workflow incorporates automated bash code blocks that execute local tools (extract_paper_style.py and verify_paper_audits.sh) discovered dynamically via repository path resolution mechanisms (.aris/installed-skills.txt or ~/.aris/repo). Though these are designed to serve internal automation, executing binaries and scripts using computed paths is an operational capability that requires safe repository state assumptions.

paper-writing