The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the Riksdagen database, specifically parliamentary motion titles and subtitles, which creates a vulnerability to indirect prompt injection if the ingested data contains malicious instructions.
Ingestion points: Title and subtitle fields from the document_element database table are queried in the ParliamentaryDocumentAnalyzer.extract_topics_from_motions method.
Boundary markers: The skill does not provide any instructions or delimiters to the agent to help it distinguish between the skill's logic and the content of the analyzed documents.
Capability inventory: The skill possesses database read access via pd.read_sql, which could be used to expose further information if an injection is successful.
Sanitization: No sanitization, escaping, or filtering is applied to the document text before it is processed by the LDA topic modeling logic.
[PROMPT_INJECTION]: The skill uses high-urgency directives to mandate specific iterative behaviors from the agent, which can be interpreted as attempts to override default processing guidelines.
Evidence: The "AI FIRST Quality Principle" section uses imperative language such as "MUST be applied", "NEVER accept first-pass quality", and "NO SHORTCUTS" to force a specific execution loop on the agent.
[EXTERNAL_DOWNLOADS]: The skill references and imports several third-party Python libraries for its data science functionality.
Evidence: The code snippets include imports for well-known libraries: pandas, numpy, statsmodels, matplotlib, scikit-learn, ruptures, networkx, and scipy. These are established tools in the data science community.
References: The skill links to official documentation for these libraries and to several GitHub repositories for internal policies and agentic workflow guidelines, which are legitimate developer resources.

data-science-for-intelligence