The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill aggregates content from project files such as README.md, package.json, and documents within the .planning/ directory, and interpolates this data into prompts for explorer and challenger sub-agents.
Ingestion points: Step 1 of SKILL.md reads .planning/PROJECT.md, .planning/ROADMAP.md, and issue descriptions.
Boundary markers: Content is placed under headers but lacks explicit safety delimiters or instructions to ignore instructions found within the ingested data.
Capability inventory: Sub-agents are defined as general-purpose, which typically include tools for shell execution and file system access.
Sanitization: No sanitization or validation of the gathered repository content is performed before injection.
[COMMAND_EXECUTION]: The skill uses shell commands and inline Node.js scripts to manage environment settings and modify configuration files.
Evidence: SKILL.md executes node -e to read from and write to ~/.claude/settings.json in order to enable experimental features.
[DYNAMIC_EXECUTION]: The script scripts/kata-lib.cjs performs file system operations that scan directories outside of the immediate project root.
Evidence: The script uses path.join(__dirname, '..', '..') to discover and read from sibling directories prefixed with kata- to resolve templates and check for configuration drift.

kata-brainstorm