Build Your AI Assistant After the Workshop

During the workshop, you created three important Context Blocks. You should receive copies of these materials by email after the workshop.

Your next step is to combine them with the source material your AI Assistant will need.

Your source material might include documents, spreadsheets, presentations, policies, process descriptions, examples, reports, reference materials, glossaries, or other information relevant to the problem.

1. Create a folder on your computer for your assistant. Give it a clear name based on the problem you are solving — for example, Customer Response Assistant.
2. Place the materials you created during the workshop into that folder: Role Context Block, Problem Context Block, and AI Instructions.
3. Identify the most useful source files for the problem. Ask yourself: "what would a capable new employee need to read before I would trust them to help me with this problem?" Those are strong candidates for your initial source material.
4. Start small. You can add additional context as you discover what the assistant needs.

Step 1 — Create a project

Open ChatGPT. Select New project from the sidebar. Give the project a name that describes the problem or assistant — for example, Customer Response Assistant.

Step 2 — Add your AI Instructions

Open the project's menu and select Project settings. Find the area for Project instructions. Paste the AI Instructions you created during the workshop into this field. These instructions tell ChatGPT how to operate whenever you work inside this project.

Step 3 — Add your Role and Problem Context

Add your Role Context Block and Problem Context Block as project sources. You can upload them as files or paste their contents into appropriate project source material.

The important distinction is: instructions tell the assistant how to work; Role and Problem Context help the assistant understand your situation.

Step 4 — Add your Context Sources

Upload the documents and other materials the assistant needs to understand and help solve the problem. Examples might include: policies, process documentation, presentations, reports, reference documents, product information, spreadsheets, examples of good previous work, and terminology or glossaries.

ChatGPT Projects can use uploaded files as persistent project context.

Step 5 — Start a new conversation inside the project

Your first conversation can be simple: "Review the instructions and sources available in this project. Summarize your understanding of my role, the problem we are trying to solve, and the most important constraints you should follow. Identify anything important that appears to be missing or unclear."

Do not immediately ask it to solve everything. First determine whether the assistant understands the environment correctly. Then begin working on the problem.

Step 1 — Create a project

Open Claude. Select Projects and create a New Project. Give it a name based on the problem you are solving — for example, Customer Response Assistant.

Step 2 — Add your AI Instructions

Select Set project instructions. Paste the AI Instructions you created during the workshop. These instructions will apply across conversations within the project.

Step 3 — Add your Role and Problem Context

Add your Role Context Block and Problem Context Block to the project's knowledge. These become part of the persistent context Claude can reference when working with you.

Step 4 — Add your Context Sources

Use the project's knowledge area to add relevant source material. Claude Projects can use uploaded documents, text files, code, and other information as persistent knowledge across project conversations.

Start with the sources most directly related to your problem. Avoid filling the project with documents simply because they might someday be useful. Relevance is more important than volume.

Step 5 — Test its understanding

Start a new conversation and ask: "Review the project instructions and knowledge available to you. Summarize your understanding of my role, the problem we are trying to solve, and the important rules or constraints. Then identify any gaps or conflicting information you see."

Correct anything that is wrong. Add missing context when necessary. Then begin using the assistant.

For many users, the closest equivalent to a persistent AI Assistant will be a Copilot agent or a Copilot Notebook. Microsoft also allows Copilot to ground responses in Microsoft 365 work information, attached files, links, and other permitted sources.

Using a Copilot agent

Step 1 — Open Microsoft Copilot

Sign into your organization's Microsoft 365 environment. Open Copilot and locate Agents. If your organization permits users to create agents, choose the option to create a new agent.

Step 2 — Add your AI Instructions

Use your workshop AI Instructions as the starting instructions for the agent. The instructions should describe: who the assistant is helping, what problem it is helping solve, what its role is, what tasks it should perform, what rules or constraints it should observe, and what kind of outputs it should produce.

Step 3 — Add your Role and Problem Context

Include your Role Context and Problem Context as part of the persistent knowledge or supporting material available to the agent. Do not assume Copilot already understands your role simply because you use Microsoft 365 — make important context explicit.

Step 4 — Connect appropriate knowledge

Depending on your organization's configuration, this may include approved Microsoft 365 content, SharePoint information, files, links, or other organizational knowledge. Copilot can ground responses using work information and content you explicitly provide, subject to your license, permissions, and administrator settings. Make sure you are pointing the assistant toward the authoritative version of important information — if several versions of the same policy, presentation, or spreadsheet exist, do not make the AI guess which one is correct.

Alternative: Microsoft Copilot notebook

If your organization gives you access to Copilot Notebooks, you can create a notebook, add relevant reference material, and provide custom instructions for how Copilot should work with that material. This can be a good starting point when your assistant is primarily intended to analyze, synthesize, or work from a defined collection of source material.

Before relying on it for important work, test whether it actually understands your context. Ask it questions such as:

• What do you understand my role to be?
• What problem are we trying to solve?
• Who is affected by this problem?
• What rules or constraints should you follow?
• Which sources would you use to answer questions about this problem?
• What information appears to be missing?
• Do any of the sources appear to contradict one another?

Then give it several real examples of work. Compare its output with what a knowledgeable person would produce. Correct errors. Add missing context. Clarify instructions. Remove outdated or contradictory sources. Then test again.

As you use the assistant, pay attention to where it struggles. When the output is wrong or disappointing, do not simply rewrite your prompt over and over. Ask:

• Is the problem clearly defined?
• Are the instructions clear?
• Is important context missing?
• Is a source outdated?
• Are two sources contradicting one another?
• Does the assistant understand our terminology?
• Have we told it the rules and boundaries?

Every failure is information you can use to improve the system. The objective is not a perfect prompt. The objective is a continuously improving AI Assistant built from clear instructions and well-curated context.