The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs several shell operations to analyze files and validate generated content.
Evidence: Uses ls -R, grep, head, and unzip for file inspection and data counting.
Evidence: Executes officecli for reading and manipulating Office documents.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from untrusted user documents to create new agent instructions, creating a vulnerability surface for indirect injection.
Ingestion points: Reads user-specified reference files via officecli read (SKILL.md Phase 1).
Boundary markers: None present to delimit user content from generated instructions.
Capability inventory: Includes file system writes, subprocess execution (bun, unzip), and officecli operations.
Sanitization: No sanitization or validation of the text content extracted from user documents is specified before it is used to write new SKILL.md files.
[DYNAMIC_EXECUTION]: The skill generates new executable instruction files and runs local scripts as part of its validation lifecycle.
Evidence: Phase 2 of SKILL.md involves writing new SKILL.md and reference files to the skills/ directory.
Evidence: The 'Validate' section instructs the agent to execute JavaScript/TypeScript files using the bun runtime (e.g., bun tests/isolated-workspace/scripts/tracer.mjs).

skill-creator