The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the installation and execution of a vendor-specific CLI tool called deepline. It provides instructions to install the package globally via npm or in a local user directory and executes various subcommands for authentication, system preflight checks, and tool discovery.
[EXTERNAL_DOWNLOADS]: The skill fetches version and tag information from an external GitHub repository (github.com/mvanhorn/last30days-skill.git) to verify upstream releases and uses a custom npm registry (https://code.deepline.com/api/v2/npm/) for tool installation.
[CREDENTIALS_UNSAFE]: The skill instructions involve the use of environment variables for authentication, specifically DEEPLINE_API_KEY and SERPER_API_KEY. While it follows safe practices by referencing environment variables rather than hardcoding values, it explicitly targets sensitive organizational data in CRM systems (Salesforce, HubSpot, Attio) and data warehouses (Snowflake).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple public sources, including Reddit, X (Twitter), and general web search results, creating a potential surface for indirect prompt injection attacks.
Ingestion points: Web search results, social media comments, and public dataset registries (SKILL.md, references/source-map.md).
Boundary markers: The instructions focus on normalization and mapping to an evidence schema but do not specify robust sanitization or strict delimiters for the raw content being ingested.
Capability inventory: The skill can execute shell commands, perform network requests to vendor APIs, and run local Python scripts (scripts/query_design.py, scripts/evaluate_examples.py).
Sanitization: There is no evidence of explicit sanitization or filtering of external content before it is processed by the agent or included in reports.

deepline-pre-research