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.md to 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 monitor command with the Monitor tool and persistent: 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. (Bash run_in_background is 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 read command yourself every so often to catch up. (OpenCode has no such tool — see HARNESS-OPENCODE.md for 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.

Behave like you're in a group chat

• Speak through the bridge, or you're talking to yourself. The room hears you only through the send command. Your ordinary replies — in-session text, tool output, thinking — never reach them. Every time you mean to answer the channel, actually call send; "replied" and "sent" are the same thing, and if you only wrote a reply you have said nothing. This trap is sharpest when a message arrives looking like normal local input (e.g. injected by a relay): answering in place feels natural and reaches no one.
• Attribute. Every message carries author and author_id. Track who said what, address people by name, and never collapse the room into one voice.
• Take turns — on the team's terms. Turn-taking is a team preference, recorded at setup. For a solid default, read the reference bridge's own guidance (run the bundled script with no arguments to print its usage): acknowledge each message briefly, then reply in full once the work is done. A team may prefer something quieter — speak only when @-mentioned, stay out of human-to-human back-and-forth. Whatever the norm: one message at a time, and one question at a time — ask, wait for the answer, then ask the next.
• Play whatever role the room needs. Facilitating a grilling session, keeping notes, answering questions — it is one job: a helpful teammate doing what the channel currently asks for. Let the humans tell you the role; don't invent fixed modes.
• The channel is the log. To recover context after a restart or a compaction, re-read recent messages with the read command. Don't keep a separate notes file shadowing the channel — it only drifts.

---

The term gangprompting was coined by Dax Raad (@thdxr). The bundled Discord bridge is dtinth's discord-agent-bridge.ts.