The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources that could contain malicious instructions designed to influence subagents.
Ingestion points: Error logs are read from the filesystem in commands/debug.md, and network/console logs are gathered via Chrome MCP tools as described in rules/chrome-evidence-tools.md.
Boundary markers: The prompts in templates/parallel-agent-prompts.md use headers like EVIDENCE: to separate data from instructions.
Capability inventory: The skill can launch parallel subagents, execute shell commands, and run JavaScript in a browser context.
Sanitization: No explicit sanitization or escaping of the ingested data is performed before it is interpolated into agent prompts.
[DYNAMIC_EXECUTION]: The skill facilitates the execution of JavaScript within the browser page context for state inspection.
Evidence: In rules/chrome-evidence-tools.md and SKILL.md, the agent is instructed to use mcp__claude-in-chrome__javascript_tool to run snippets like JSON.stringify(localStorage) or access React internals.
[COMMAND_EXECUTION]: The skill runs local shell commands to retrieve project context and logs.
Evidence: The /debug command in commands/debug.md executes tail -50 logs/error.log, git log, and git diff.
[METADATA_POISONING]: The skill contains conflicting author and repository information in its metadata files.
Evidence: The .claude-plugin/plugin.json file identifies the author as "Jeremy Dawes" and the repository as a jezweb GitHub account, which contradicts the provided skill author context of "fefogarcia".

deep-debug