The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The transcript analysis service logic in refs/analysis-patterns.md contains a vulnerability surface where untrusted data could influence agent behavior.
Ingestion points: Transcript content is retrieved from the Supabase transcripts database table and interpolated into an AI prompt.
Boundary markers: The prompt template lacks delimiters (such as XML tags or triple quotes) or specific instructions to the model to treat the transcript content as data rather than instructions.
Capability inventory: The AI's response is used to update the themes, key_quotes, and ai_summary fields in the database via the supabase.from('transcripts').update method.
Sanitization: The implementation uses JSON.parse on the raw AI response without validating the data against a schema or sanitizing the content before the database write operation.

data-analysis