The Story of Creating an AI Meeting Room on Discord and Having Four Roles Evaluate Issues

I created an AI meeting room on Discord that allows a single topic to be evaluated by multiple roles.

Initially, the setup was just for the AI to answer questions when asked by a bot. However, for complex design consultations, the first answer is not always the correct one.

Therefore, instead of having the same AI answer everything at once, I separated the roles and the order.

Four-Stage Meeting

Post the topic
↓
Planner: Organize objectives, premises, options, and recommended policies
↓
Critic: Point out oversights, risks, and opposing opinions
↓
Dev: Concrete implementation plan based on the feedback
↓
Minutes: Document decisions and pending items

The results of each stage are posted to Discord in order. Not only the final conclusion but also the concerns raised along the way are preserved.

More than 'Multiple AIs', it's 'Multiple Roles'

What was important was not the number of models.

It is about explicitly assigning roles: one to plan, one to question, and one to implement. A Planner alone might move forward too optimistically. A Critic alone might stop at pointing out flaws.

By setting an order, criticism becomes material for the next improvement.

Why I chose Discord as the entry point

I thought it would be easier to operate by placing it in Discord, which we use daily, rather than creating a new dedicated screen.

In addition to slash commands, I made it possible to consult Claude Code from regular text in designated meeting rooms. However, I prevent reply loops by not letting the bot react to its own posts.

In the forum format, it also checks the parent channel ID. Since threads are separated by topic, it has become easier to trace the history later.

Stop if it fails midway

Since the AI is called at each stage, the meeting can take several minutes.

If a call in the middle fails, forcing the rest to continue would result in meeting minutes missing the necessary premises. Therefore, it stops at the stage where the error occurred and returns the status to Discord.

It is designed to stop if the premise is broken, rather than forcing it to the end just because it is automated.

What I learned from creating an AI meeting room is that the key to getting good answers is not just long prompts. When you design a conversation that includes roles, order, stop conditions, and recording locations, it becomes easier to connect AI answers to decision-making.

#AIAgent #DiscordBot #ClaudeCode #PersonalDevelopment #GenerativeAI #SystemDesign #BusinessImprovement