The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates a Node.js validation script at runtime and executes it to verify the integrity of the knowledge graph.
Evidence: In SKILL.md (Phase 6), the agent is instructed to write a specific JavaScript template to $PROJECT_ROOT/.understand-anything/tmp/ua-inline-validate.cjs and then execute it using node.
Evidence: In SKILL.md (Phase 0.5), a Node.js one-liner is executed via node -e to generate the .understandignore file based on project structure.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external data (the user's codebase) which could contain malicious instructions designed to influence the agent's behavior during analysis.
Ingestion points: The skill reads all project files during Phase 2 (Analyze) and processes manifests/READMEs in Phases 1, 4, and 5.
Boundary markers: The instructions use markdown code blocks as delimiters when passing file content to subagents, but these can be bypassed by adversarial content.
Capability inventory: The skill has the capability to write files, execute local scripts (Node/Python), and dispatch further subagents with project context.
Sanitization: No explicit sanitization or escaping of file content is performed before interpolation into subagent prompts.
[COMMAND_EXECUTION]: The skill executes shell commands to manage its own development environment and dependencies.
Evidence: In SKILL.md (Phase 0), the script runs pnpm install and pnpm --filter @understand-anything/core build to ensure the core library is available before analysis begins.

understand