The Agent Skills Directory

[PROMPT_INJECTION]: Obfuscation via homoglyphs was detected in evidence.json. The identifier verdict-matе (line 35) uses a Cyrillic 'е' (U+0435) instead of the Latin 'e' (U+0065). This is a common technique used to evade text-based security filters or cause logic errors during data processing.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its design, which involves ingesting and acting upon untrusted data from users and external web sources.
Ingestion points: Processes claims provided by the user and retrieves content from external websites or documents (Step 1 and Step 3 in SKILL.md).
Boundary markers: The instructions provide a specific workflow to isolate atomic assertions and require a structured output template for the final verdict to help maintain agent context.
Capability inventory: The skill allows the agent to fetch arbitrary URLs and execute several MCP tools (e.g., budgetkey, knesset) to retrieve data.
Sanitization: The instructions explicitly direct the agent to strip rhetoric and opinion from input and strictly forbid the fabrication of figures not found in the source material.
[EXTERNAL_DOWNLOADS]: The skill fetches data and configuration from well-known Israeli government portals and official financial services, including cbs.gov.il, boi.org.il, next.obudget.org, and data.gov.il.

israeli-fact-checker