The Agent Skills Directory

[PROMPT_INJECTION]: No override, bypass, or 'ignore previous instructions' patterns were detected. The instructions use strong directive language (e.g., 'Always read', 'Read... when available') but these are purely operational and aligned with the skill's planning purpose.\n- [DATA_EXFILTRATION]: No evidence of exfiltration. While the skill has access to tools like WebFetch and Bash, the instructions are limited to reading local project files (e.g., main.tex, memory/*.md) and writing the contract to the .agent/ subdirectory. No sensitive system paths (e.g., ~/.ssh, ~/.aws) are accessed.\n- [REMOTE_CODE_EXECUTION]: No remote scripts are downloaded or executed. The skill does not use npm, pip, or any other package manager to install external dependencies at runtime.\n- [COMMAND_EXECUTION]: The skill uses the Bash tool for project navigation and file reading. There are no patterns of arbitrary command execution or unsafe shell interpolation of user input.\n- [CREDENTIALS_UNSAFE]: No hardcoded API keys, tokens, or secrets were found. The skill does not seek to read .env files or system credential stores.\n- [INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it reads untrusted external data (the paper draft files and project memory). However, its operations are restricted to structured planning, and it lacks the high-privilege capabilities (like network exfiltration of read data) that would elevate this surface beyond a low-risk category.

paper-writing-contract-planner