The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied context or external source material (via the --source option) to generate interactive HTML explainers. This creates an attack surface where malicious instructions embedded in the source material could potentially influence the agent's behavior during the generation process.
Ingestion points: User conversation context and designated external files, issues, or URLs.
Boundary markers: Implements provenance marking using specific spans (SOURCE-ESTABLISHED, INFERENCE, OPEN) to distinguish evidence types.
Capability inventory: Ability to write HTML files to the local file system (defaulting to ~/Documents/explainers/) and execute verification scripts.
Sanitization: Generates a restrictive Content Security Policy (CSP) for the resulting HTML files and enforces subresource integrity for external dependencies like MathJax.
[COMMAND_EXECUTION]: The skill includes several Python scripts (verify_explainer.py, verify_rendered.py, mutate_gate.py) that are designed to be executed by the agent to validate output artifacts. These scripts use subprocess.run to manage the execution of other scripts and tool installation.
Evidence: mutate_gate.py executes other scripts via uv run --script. verify_rendered.py uses the shell to perform environment setup.
[EXTERNAL_DOWNLOADS]: The verify_rendered.py script automatically downloads and installs the Chromium browser using the Playwright framework to enable headless rendering tests.
Evidence: The script calls playwright install chromium via subprocess.run. This is a download from a well-known service (Microsoft Playwright).

teachback