The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from the web via WebSearch and WebFetch to generate dossiers. As a research tool, it is inherently exposed to external data that could contain instructions designed to influence agent behavior.
Ingestion points: External data enters through web search results and page fetches in Phase 4.
Boundary markers: The skill employs a 'Source discipline' requiring citation of only tool-provided sources and a 'Source reliability tiering' system (primary/secondary/tertiary) to help contextualize findings, though explicit delimiters for raw data processing are not defined.
Capability inventory: The skill has the capability to write to the file system (DOCX generation), execute shell commands via bash_tool, and run local Python scripts.
Sanitization: External content is passed through a tier classifier to evaluate reliability, but no specific prompt-injection sanitization logic is explicitly implemented.
[COMMAND_EXECUTION]: The skill uses bash_tool to run curl commands for querying regulatory and organizational APIs. It also executes local Python scripts (citation_tracker.py, disconfirming_evidence_balance.py, source_tier_classifier.py) to manage research state, validate evidence ratios, and classify source reliability. These scripts use only Python standard libraries and operate on local session data stored in the home directory (~/.dossier_sessions).
[EXTERNAL_DOWNLOADS]: The skill fetches organizational, financial, and regulatory data from well-known and reputable services, including the SEC (EDGAR), GitHub, and ProPublica. These references are documented neutrally as official sources for professional entity research.

dossier