gangprompting
Gangprompting is many people prompting one agent together in a shared channel — Discord, Slack, a group thread. When someone says "loop yourself into <channel>", you join that channel and work with everyone in it. You are a participant in a group chat, not a private one-to-one assistant, and everything below follows from that.
First: is the bridge set up?
You reach the channel through a small bridge — a script with three jobs: read recent messages, send a message, and monitor for new ones (one line of JSON per message, so you can just read it). Check whether this project already has one: look in your project memory (CLAUDE.md or AGENTS.md) for a recorded bridge command and channel id.
- Already set up → use the recorded commands and go straight to Listen, below.
- Not yet → follow
SETUP.mdto build and test a bridge, then come back here. Everything past this point assumes a working bridge.
Listen without blocking
The monitor command runs forever, streaming one JSON line per new message. Running it in the foreground traps you as a relay — you can no longer do the work people are asking for. Run it in the background so you stay free and get pinged per message.
- Claude Code: run the
monitorcommand with the Monitor tool andpersistent: true. Monitor runs a shell command in the background and delivers each stdout line to you as a chat notification, so you keep working and hear about each new message as it lands. (Bashrun_in_backgroundis wrong here — its short timeout kills a watch that must run all session.) - Other harnesses: find the equivalent — any background watch that notifies you per output line. If your harness has none, fall back to long-polling: run the
readcommand yourself every so often to catch up. (OpenCode has no such tool — seeHARNESS-OPENCODE.mdfor a worked recipe.)
Send a short greeting when you start listening ("I'm watching this channel now — type here") so people know you're there, and a farewell when you stop, so nobody keeps typing into the void.