The Agent Skills Directory

[DATA_EXFILTRATION]: The skill's orchestration logic explicitly references accessing configuration files at ~/.config/claude-seo/google-api.json to share Google API credentials between skills. Accessing sensitive configuration data within the user's home directory is a high-severity data exposure risk.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to automatically ingest untrusted data from the project root (e.g., BRAND.md, VOICE.md, DISCOURSE.md) and external web search results. This content is then fed into downstream agents (blog-writer, blog-reviewer) that have capabilities to write files and execute subprocesses. Although the skill implements a 'nonce + fence' mechanism using a helper script (load_untrusted_root.py) to mitigate injection, this defense relies on the agent's consistent adherence to a complex set of instructions while processing adversarial data.
[COMMAND_EXECUTION]: The skill utilizes a 'Delivery Contract' that automatically executes multiple local scripts (scripts/blog_preflight.py, scripts/blog_render.py, scripts/generate_hero.py, scripts/lint_prose.py) to generate artifacts, render HTML, and perform audits. This automated loop creates a significant execution surface that processes data derived from untrusted external sources.
[EXTERNAL_DOWNLOADS]: The skill connects to various remote services, including Google APIs, Pixabay, Unsplash, and Pexels, for research and media generation. While these are well-known technology services, the aggregate level of network activity and the flow of this data into an automated execution pipeline increases the system's attack surface.

blog