The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a pattern where untrusted data (user messages) is ingested into an LLM context, creating a surface for indirect prompt injection.
Ingestion points: The loadContext function in AGENTS.md (and the convex/index.ts example) fetches message data from the messages table, which includes user-contributed content.
Boundary markers: Absent. The code builds an array of messages and passes it directly to the OpenAI API without using specific delimiters or instructions to the LLM to ignore embedded commands.
Capability inventory: The skill possesses capabilities to modify the database state using ctx.db.insert and ctx.runMutation (via the writeAgentResponse function) and schedule future tasks.
Sanitization: Absent. There is no evidence of filtering, escaping, or validation of the message content before interpolation into the LLM prompt.
[EXTERNAL_DOWNLOADS]: The skill's documentation includes a package.json file referencing official Node.js packages for Convex and OpenAI.
Packages: convex@^1.31.2 and openai@^4.79.0 are standard libraries for interacting with their respective services.

convex-backend