The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local repository and interpolates it into a generated HTML report, creating a potential vector for malicious content to influence the user's browser session. \n- Ingestion points: The skill processes codebase commit history via git log, reads domain documentation in GLOSSARY.md, architectural records in docs/adr/, and performs an organic walk of all source files in the project. \n- Boundary markers: The instructions lack specific guidance for using delimiters or markers when interpolating repository content into the HTML template. \n- Capability inventory: The skill can write to files (GLOSSARY.md), execute shell commands for git operations and opening files (open, xdg-open, start), and create new files in the system temporary directory. \n- Sanitization: No sanitization or escaping of repository-sourced text is specified before it is embedded in the report template. \n- [COMMAND_EXECUTION]: The skill uses platform-specific system commands (xdg-open, open, start) to automatically launch the generated architectural report in the user's default browser. This execution path uses paths derived from environment variables like $TMPDIR. \n- [EXTERNAL_DOWNLOADS]: The generated HTML reports fetch CSS and JavaScript assets from established, well-known content delivery networks including Tailwind CSS and jsDelivr (for Mermaid.js). These are used to provide styling and diagramming functionality for the report. \n- [DYNAMIC_EXECUTION]: The HTML report template configures the Mermaid diagramming engine with its securityLevel set to loose. This configuration allows for the rendering of HTML tags and execution of scripts within diagrams, which increases the potential impact of an indirect prompt injection attack if the repository contains malicious diagram syntax.

improve-codebase-architecture