The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a pattern for indirect prompt injection by design. It ingests untrusted data from external chat platforms (Discord/Slack) and injects it into the agent's prompt context.
Ingestion points: Messages are streamed into the session via the Monitor tool (Claude Code) or opencode-relay.ts (OpenCode).
Boundary markers: opencode-relay.ts prefixes injected messages with a source label (e.g., "From Discord, channel [ID]:"). However, there are no instructions to the agent to treat this content as untrusted or to ignore embedded commands.
Capability inventory: The agent uses discord-agent-bridge.ts to read, send, edit, and delete messages, in addition to its standard toolset.
Sanitization: External content is encapsulated in a JSON record before injection, but the text content of the messages is not sanitized or filtered for instructions.
[REMOTE_CODE_EXECUTION]: SETUP.md and the header of discord-agent-bridge.ts suggest running the bridge script directly from a remote URL using Deno.
Evidence: deno run --allow-net --allow-env --allow-read https://raw.githubusercontent.com/dtinth/discord-message-proxy/main/tools/agent-bridge/discord-agent-bridge.ts.
Note: The remote resource is hosted on the author's own GitHub repository (dtinth), which aligns with the expected vendor infrastructure for this skill.

gangprompting