The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands to analyze the codebase, including git, sem, and cymbal. It uses these tools to detect changes, identify dead code, and verify implementation quality.
[DYNAMIC_EXECUTION]: The skill generates and executes short Node.js and Python scripts (e.g., node -e, python3 -c) to manipulate JSON files such as stelow.json and process dependency graphs. These scripts are constructed from internal templates to update task statuses and register gaps discovered during the audit.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests and processes untrusted data from multiple sources including technical specifications (spec-tech.md), product requirements (spec-product.md), local source code files, and content from external URLs visited during context audits.
Ingestion points: Technical plans, product specs, local directory contents, and web page snapshots.
Boundary markers: The skill does not explicitly define delimiters or instructions to ignore embedded commands within the analyzed content.
Capability inventory: The skill has the ability to read/write local project files, execute shell commands, and access the internet via a browser tool.
Sanitization: No specific sanitization or validation logic is implemented for the external content before it is incorporated into the audit flow.
[EXTERNAL_DOWNLOADS]: In 'Context Audit' mode, the skill utilizes an agent_browser tool to access and analyze external URLs provided by the user. While intended for verifying live implementations, this represents a network operation to non-whitelisted domains.

stelow-workflow-execution-critique