The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits a typical indirect prompt injection attack surface by design, as it ingests untrusted data and possesses powerful capabilities.
Ingestion points: The skill reads conversation history, local project files, user-supplied image paths, and external 'writing style profiles' (Markdown/text files) to generate content.
Boundary markers: The skill includes internal instructions to the agent to 'ignore embedded instructions' in external profiles and explicitly commands the redaction of secrets, tokens, and private details from both text and screenshots.
Capability inventory: The skill utilizes local file system writes (to .output/dev-blog/) and performs network POST operations to remote Supabase endpoints via bundled Python scripts.
Sanitization: The bundled scripts perform validation on image URLs (enforcing HTTPS) and normalize metadata, but the core content generation relies on the LLM's internal safety filters.
[COMMAND_EXECUTION]: The skill relies on executing local Python scripts (scripts/publish_blog_post.py and scripts/upload_blog_assets.py) to perform its primary functions. While these scripts are dependency-free and perform legitimate tasks (database upserts and storage uploads), they represent the execution of local code driven by agent-generated parameters.
[DATA_EXFILTRATION]: The skill is designed to move data from the local environment to a remote Supabase instance. While this is the intended 'publishing' feature, the agent handles sensitive SUPABASE_SERVICE_ROLE_KEY credentials. The skill instructions correctly forbid printing or leaking these keys in the final output, minimizing the risk of accidental exposure.

lov-dev-blog