The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill scans the repository for code concepts and uses the discovered content to generate questions for automated validation sessions via the claude CLI. This creates a surface for indirect prompt injection where malicious strings within the analyzed code could influence the behavior of the validation agent. * Ingestion points: Phase 1 (Discovery) scans repository entry points, API routes, and configuration files. * Boundary markers: The validation prompt instructions require the agent to use only repository documentation, but the prompt construction does not include strong delimiters or sanitization for the injected concept data. * Capability inventory: The skill has the ability to read all files, execute bash commands, and commit changes to the repository. * Sanitization: No sanitization or escaping is performed on code-extracted strings before they are interpolated into the validation prompts.
[DATA_EXFILTRATION]: The skill extracts environment variables, API routes, and database schemas to generate documentation files like docs/generated/env-vars.md. This involves reading files such as .env.example and code patterns like process.env. While intended for documentation, this process can lead to the exposure of sensitive configuration if the repository contains secrets that are not properly excluded.
[DYNAMIC_EXECUTION]: The skill uses python3 -c to generate JSON-formatted knowledge maps in the /tmp directory at runtime. It also constructs and executes shell commands that incorporate variables derived from the repository analysis, such as repository names and concept details.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to perform repository operations, including git commands for committing documentation updates, filesystem operations for file management, and invoking the claude CLI for documentation validation.
[EXTERNAL_DOWNLOADS]: The skill references several companion tools and its own installation path via npx from the fellowship-dev organization. These are vendor-provided resources associated with the skill author's own infrastructure.

popsicle