The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Python's subprocess module in several utility scripts (run_eval.py, improve_description.py, generate_review.py) to execute the platform's own claude CLI for testing and optimization purposes, and the lsof utility to manage local server ports. These operations are performed locally within the developer's environment and are necessary for the skill's core functionality.
[DYNAMIC_EXECUTION]: As a meta-development tool, the skill dynamically generates SKILL.md files, scripts, and configuration data based on user templates and iteration feedback. It facilitates the execution of these generated components via subagents or local CLI calls to measure triggering accuracy and performance.
[INDIRECT_PROMPT_INJECTION]: The skill's iteration and optimization phases involve processing external data, such as user-provided test prompts (evals.json) and evaluation feedback (feedback.json). These inputs are used to refine the instructions of the skill being developed. The skill implements structural countermeasures, such as commitment gates and explicit behavioral contracts (e.g., in the 'Designing for Agent Cognition' section of SKILL.md), to maintain instruction following and mitigate risks associated with untrusted data ingestion.
[OBFUSCATION]: The generate_review.py script and viewer.html use Base64 encoding to embed binary content (images, spreadsheets, and PDF previews) directly into generated HTML reports. This is a standard technique for creating self-contained local visualization files and is not used to hide malicious activity or bypass security scans.

custom-skill-creator