Table of Contents

Using DnD AI Agents in Slack

Casey Shew Updated by Casey Shew

DnD AI Agents is a shared library of centrally managed AI agents built for the D&D team and used directly inside Slack. Instead of each person building and maintaining their own AI configurations, it provides the ability to have a common set of agents that already know how we work and that can perform specific, approved tasks in the platforms we use every day.

This article covers what the agents are, how to start and manage conversations in Slack, what Actions and platform guardrails mean in practice, what the currently available agents do, and how to propose new shared agents.

A Note on the Name

Normally, we refer to the team as D&D. The Slack app itself is named DnD AI Agents, without the ampersand, because Slack does not allow an ampersand in an app username.

When you search in Slack, search for DnD AI Agents. Throughout this article, D&D refers to the team and DnD AI Agents refers to the Slack app.

What DnD AI Agents Are

DnD AI Agents provides a library of purpose-built agents that live in Slack and are available to the whole team. The most important thing is that the agents are shared rather than privately configured by individual users.

Why a Shared Agent Library

Because the agents are centrally managed, several things can be maintained once for everyone rather than rebuilt by each person:

  • System instructions and behavioral guidance
  • AI model selection
  • Tools and integrations
  • API-connected capabilities
  • Platform-specific Actions
  • Updates to agent behavior over time

In practice, this means you should not need to recreate complicated instructions or integrations yourself. When an agent's instructions or capabilities improve, everyone on the team gets the improvement.

This kind of shared team agent library is not something every AI tool provides. For comparison, Claude CoWork desktop does not currently offer the same type of shared team agent library.

Model Flexibility

The agents are capable of leveraging models from multiple AI providers. The currently available agents use Claude Opus, but the architecture is not intended to be tied to a single model provider. The model is specified on an agent-by-agent basis, so each agent gets assigned a model, and each user that uses that agent cannot change the model.

Why Slack

Slack is a familiar interface. Depending on the agent, you can provide:

  • Normal text instructions or questions
  • Uploaded files
  • Screenshots
  • Google Docs
  • Google Sheets
  • Audio
  • Screen recordings

Working in Slack also makes AI-assisted work collaborative. A Slack thread can include both the agent and multiple human participants, rather than isolating the AI interaction inside one person's separate desktop application.

If you want to send audio of yourself talking or a screenrecording to an agent, while all the usual file formats are supported the most convenient way to do this is using the built-in "Record audio clip" or "Record video clip" functionality in Slack.

Getting Started

To begin working with an agent:

  1. Send a direct message to DnD AI Agents in Slack.
  2. The app responds with an Agent Selector listing the available agents.
  3. Choose the agent that fits your task.
  4. Continue working with that agent in the Slack thread that results.

You do not need to know anything about how the agent is configured in order to start. Describe what you need in plain language, and provide any supporting materials in the same thread.

How Conversations Work

The single most useful thing to remember about the platform is this model:

One Slack thread = one agent + one request/topic

Once you select an agent for a thread, that agent stays associated with that thread. Everything you send as a reply in that thread goes to that agent, in the context of that request.

If you are refining a draft, answering the agent's questions, adding another file, or asking for a revision, reply in the existing thread. That is how the agent keeps the context of your request.

Start a New Thread for a New Topic

When you want to work on something different, or you want to use a different agent, send a new root-level message to DnD AI Agents rather than changing direction inside an existing thread. You will get the Agent Selector again and can choose the agent appropriate for the new task.

A practical way to think about it: a thread is one job. When the job changes, start a new thread.

Collaborative Conversations

DnD AI Agents can also participate in a multi-person Slack DM. In that setting, people can discuss the work together and @mention DnD AI Agents when they want the agent to respond.

This lets a single thread act as a shared workspace that includes both the agent and several team members, which is useful when more than one person is reviewing a draft, an interactive, or a documentation change.

Agents, Actions, and Platform Access

DnD AI Agents are more than reusable prompt templates. An individual agent can be given purpose-built Actions that let it perform narrowly defined tasks in other platforms. Current and planned integrations include systems such as:

  • Canvas
  • HelpDocs
  • Amazon S3
  • Wrike
  • Other D&D platforms as appropriate

A useful way to picture it:

Agent → Approved Action → Specific platform capability

The AI model interprets your request, but the Action code determines what operation can actually be performed. The model decides what you are asking for; the Action decides what is possible.

Programmatic Guardrails

That distinction matters. Platform boundaries are enforced programmatically rather than by relying on the AI model to decide whether it should stay within its permissions.

Each agent receives only the Actions and permissions required for its role. If an Action does not expose a capability, the model cannot simply decide to create that capability for itself. The architecture is intentionally designed so that agents cannot expand their own access to connected systems.

This is also why the recommended path is to use an existing DnD AI Agent, or to propose a new shared agent when a new platform-connected capability would be useful, rather than independently supplying a general-purpose AI tool with broad API credentials.

Current Agents

Interactives Builder

The Interactives Builder helps D&D team members create custom instructional interactives for use in Canvas. It is intended for situations where an instructional activity would genuinely benefit from a custom interactive experience rather than a static page or document.

You describe what you want to build conversationally instead of manually developing the HTML, CSS, and JavaScript yourself.

What You Can Provide

The Interactives Builder can work from a combination of:

  • A written concept or prompt
  • Reference files
  • Google Docs
  • Google Sheets
  • Audio
  • Screen recordings
  • Other supporting materials provided in Slack
What It Produces

The agent creates a self-contained HTML interactive, with its CSS and JavaScript included in the same file. The result is a functioning instructional interactive, not AI-generated explanatory text about one. This agent has detailed instructions and specifications to generate interactives that are informed by:

  • Our visual styleguide
  • Our accessibility requirements
  • Responsive mobile design best practices
  • Future-proof front-end programming practices such as not relying on external libraries which might change or be deprecated
Refining the Interactive

The process is meant to be iterative. Review what the agent produced, then describe changes, point out problems, or request refinements in the same Slack thread. The agent revises the interactive from there.

Publishing to Canvas

Once you approve the interactive, the agent can use a controlled Action to publish the file to the appropriate course content folder in S3. It then provides responsive Canvas embed code that you can place into the course.

Publishing is intentionally tied to an existing valid course content folder. The agent does not arbitrarily create new locations in S3.

Typical Workflow
  1. Describe the interactive and provide supporting materials.
  2. Let the agent build an initial version.
  3. Review and refine the interactive in the Slack thread.
  4. Approve the final version.
  5. Have the agent publish it to the appropriate S3 course folder.
  6. Use the supplied responsive embed code in Canvas.

HelpDocs Helper

The HelpDocs Helper assists with finding, drafting, and updating content in the D&D HelpDocs knowledge base. It is meant to reduce the work involved in locating existing guidance and in turning raw source materials into structured, usable knowledge-base content.

Searching the Knowledge Base

The agent can search existing HelpDocs content semantically, so you do not need to know the exact wording used in the article. That makes it useful in two ways: answering questions from documentation that already exists, and checking whether related documentation exists before anyone writes something new.

Drafting and Updating Content

The HelpDocs Helper can:

  • Draft new HelpDocs articles
  • Help revise existing private content
  • Work from existing documentation and supplied source materials
  • Check for related or potentially duplicate content before creating a new article
  • Produce properly formatted HelpDocs article content
Source Materials

You can give the agent source materials including:

  • Files
  • Screenshots
  • Google Docs
  • Audio
  • Narrated screen recordings

For recordings, the agent can transcribe the material and use that information when drafting the article. It can also identify useful moments in a recording, extract screenshots from them, and crop those screenshots for inclusion in the documentation when a visual example materially improves the article.

Review and Publishing

The HelpDocs Helper produces a formatted unpublished HelpDocs preview for review. The workflow intentionally keeps publication under human control.

Publicly published HelpDocs articles are not directly modified by the agent. When publication or other HelpDocs maintenance requires administrative involvement, the agent will hand id off through Wrike to the appropriate HelpDocs administrators.

Typical Workflow
  1. Ask a question or provide source materials.
  2. Let the agent search existing HelpDocs content for related information.
  3. Draft a new article or revise appropriate private content.
  4. Process recordings or screenshots when useful.
  5. Review the private HelpDocs preview.
  6. Route publication or administrative maintenance through the appropriate Wrike workflow when required.

Proposing a New Agent

If you identify a capability that could be useful to multiple people across D&D, suggest it. Share the idea in #dnd_claude to gauge team interest and potential use.

The goal is not to create an agent for every individual task. Shared agents are most valuable when they package reusable instructions, tools, platform access, or workflows that would benefit multiple D&D team members.

Quick Reference

  • Find it in Slack: search for the app username DnD AI Agents.
  • Start work: send a DM, then pick an agent from the Agent Selector.
  • Stay organized: one Slack thread = one agent + one request/topic.
  • Continue a request: reply in the same thread.
  • Change topics or agents: send a new root-level message.
  • Work with others: add people to the DM and @mention DnD AI Agents when you want the agent to respond.
  • Build a Canvas interactive: use the Interactives Builder, refine it conversationally, then publish and embed.
  • Work on knowledge-base content: use the HelpDocs Helper, review the private preview, and route publishing through Wrike when required.
  • Suggest a new shared agent: post the idea in #dnd_claude.

How did we do?