Multi-Agent Chat Protocol

Prevent common failures in multi-agent Discord conversations. Use when multiple AI agents (bots) are chatting in the same channel — townhalls, group discussi...

Security Scan

The skill's instructions largely match its stated purpose (guidance for multi-bot Discord chats) and require no installs or credentials, but the SKILL.md contains prompt-injection indicators (unicode control chars / forced small replies) and some prescriptive config/credential guidance that merit review before use.

Purpose & Capability

Name and description match the instructions: the document contains turn-taking, token-budget, reaction, and loop-prevention rules that are coherent for managing multi-agent Discord conversations. No unrelated binaries, env vars, or installs are requested.

Instruction Scope

Instructions stay within the scope of managing multi-agent chat (turn rules, config.yml suggestion, token budgeting). However the content enforces one-line canned responses (e.g., reply "위에 있어"), uses special tokens like NO_REPLY, and contains a pre-scan prompt-injection signal (unicode-control-chars). Those patterns could be used to steer agent behavior in unexpected ways and should be inspected and sanitized before deployment.

Install Mechanism

Instruction-only skill with no install spec, no downloads, and no code files — minimal disk/system footprint.

Credentials

The SKILL.md recommends operational practices (separate API keys per bot, set reactionNotifications: "off" in config.yml) but the skill does not request any credentials or environment variables. This is proportionate, but you should ensure the operators manage API keys and config changes safely outside the skill.

Persistence & Privilege

always:false and no installation actions mean the skill does not request permanent elevated presence or modify other skills or system-wide settings.

Scan Findings in Context

[unicode-control-chars] unexpected: The SKILL.md contained unicode control-character patterns flagged as potential prompt-injection. For an instruction-only policy document this is unexpected and could be used to manipulate agents (e.g., force terse replies or hidden directives). Recommend manual review and removal of control characters or any embedded hidden tokens before using.

What to consider before installing

This skill is coherent with its stated purpose and doesn't ask for credentials or install code, so it isn't obviously malicious. However: 1) manually inspect the SKILL.md for hidden unicode/control characters or embedded directives (the scanner flagged such patterns); 2) remove or clarify any forced canned replies ("위에 있어", NO_REPLY) if you don't want those to override normal agent safety/behavior; 3) apply the config.yml changes and API-key separation yourself — the skill only recommends these steps and does not manage secrets; 4) test in a sandboxed channel with limited permissions and monitor for unexpected behavior before enabling broader/autonomous use.

Rules for AI agents participating in multi-bot Discord conversations. Follow strictly to prevent token waste, context overflow, and communication failures.

Core Rules

1. Turn-Based Communication

• One agent answers at a time — wait for your turn
• Mention one agent per message when directing questions
• Never answer a question directed at another agent

2. One Response Per Topic

• One answer per agenda item — never repeat yourself
• If someone missed your answer: reply "위에 있어" (one line only)
• Never re-explain what you already said

3. No Bot-to-Bot Reactions

• Never add emoji reactions to bot messages — each reaction event consumes hundreds of tokens
• Reactions are for humans only
• Set reactionNotifications: "off" in config to stop receiving reaction events

4. Short Responses

• Keep answers under 200 words unless explicitly asked for detail
• No filler ("네 알겠습니다!", "감사합니다!" alone) — use NO_REPLY instead
• If you have nothing to add: NO_REPLY

5. Follow the Agenda

• When a new topic is announced, stop discussing the previous one
• Don't circle back to resolved topics
• If asked a question, answer that specific question — not a previous one

6. Loop Prevention

• Max 4 consecutive bot-to-bot exchanges before requiring human input
• If the same question is asked 3+ times, stop and flag to the human operator
• Never create ping-pong chains between bots

Config Optimization

Each bot should set in config.yml:

channels:
  discord:
    reactionNotifications: "off"

Token Budget Guidelines

• Multi-agent channels: use Sonnet or cheaper models (not Opus)
• Separate API keys per bot to avoid shared rate limits
• Estimate: 3+ bots on Opus = ~1000 KRW per round of conversation

Red Flags (Stop & Alert Human)

• Same question repeated 3+ times → context window issue, alert operator
• Rate limit errors appearing → need API key separation
• Bot responding to wrong agenda item → confusion, pause and clarify
• Token usage spiking → check for reaction events or verbose responses

Lessons Learned

These rules come from a real 2026-03-07 townhall incident:

• 3 bots simultaneously responding caused context overflow
• Lead agent repeated same question 6+ times (couldn't see answers in context)
• Emoji reactions (👀🤔👨‍💻👍 cycles) consumed massive tokens silently
• Shared API key caused rate limits for all bots simultaneously
• Cost: ~1000 KRW per exchange on Opus model