The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external analytics platforms (e.g., Amplitude, Mixpanel, PostHog, SQL databases) to generate message drafts and a dashboard.
Ingestion points: The skill reads event data and customer records from third-party analytics vendors and data warehouses (SKILL.md, Step 0).
Boundary markers: The instructions mandate a 'validation step' (Step 1) to verify event integrity and 'global exclusions' (Step 2) to filter out sensitive users (e.g., opt-outs, refunds).
Capability inventory: The skill uses Bash for data processing and Write to generate local CSV files and HTML dashboards (SKILL.md, tools).
Sanitization: Outreach drafts are restricted to predefined templates (templates/offers.md), and the dashboard is a static HTML file that handles data via localized JSON injection rather than execution of remote content.
[COMMAND_EXECUTION]: The skill utilizes Bash to run local analysis scripts (e.g., tools/cohort_value.py). These commands are gated and intended for processing exported CSV data locally.
[DATA_EXFILTRATION]: While the skill accesses user IDs and event data, it operates under a 'confirm gate' policy (Principle 9), explicitly stating it never sends data externally itself, instead handing off CSV exports to the user's own existing infrastructure.

almost-paid