The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents patterns for building AI agents that ingest untrusted user input and external data (via RAG).
Ingestion points: The userMessage argument and retrieved document content (relevantDocs) in convex/ai.ts (references/agents.md) are processed by LLM actions.
Boundary markers: The ragChat pattern uses system prompt instructions to delimit context, though it lacks explicit "ignore embedded instructions" warnings for the context body.
Capability inventory: The agent patterns include capabilities to perform database writes (ctx.db.patch, ctx.db.insert) and call external tools or search actions (ctx.runAction) in convex/ai.ts.
Sanitization: The skill utilizes Convex's built-in v validators for all input arguments and employs JSON.parse for handling tool-calling arguments, which provides a layer of structural validation.
[EXTERNAL_DOWNLOADS]: The skill references several official and well-known dependencies for its functionality.
Documentation suggests installing official Convex components: @convex-dev/agent, @convex-dev/rate-limiter, @convex-dev/action-retrier, @convex-dev/migrations, @convex-dev/workpool, and @convex-dev/eslint-plugin.
It also references well-known AI libraries: ai and openai.
Code examples show communication with well-known services: api.stripe.com, api.resend.com, and api.openai.com.

convex