Automate Your Weekly Reports With This AI Agent (Claude + monday.com)

What Data Does An AI Agent Need To Write Your Reports?

Whatever you want in the report has to be tracked in monday.com first, because the agent can only report on data it can see. Tracking it in even the most basic form is enough.

In our work with clients, we often see people preparing weekly, monthly or quarterly reports spend a lot of the time just finding the data, then a lot more reshaping it into the report. The agent takes over the reshaping, as long as the data is already in the system.

Our demo projects board holds new, active, stuck and completed projects, with type, expected completion, priority, revenue, cost and gross profit, plus subitems with their own status, due date and task type. We used one board, but the agent can pull from as many as you need.

What Does The Weekly Report Agent Do?

It's built to read the projects board every Monday morning, write up how each project is going and email the report as a PDF. A PowerPoint would work too, and we let the AI decide the layout.

We built it in the Claude Console at platform.claude.com rather than inside monday.com. As an admin, go to the Managed Agents area on the left and press the plus button to start a new agent.

If you already have a report and know which data points it needs, the instructions are easier to write.

How Do You Get Claude To Write The Agent's Instructions?

Dictate the job into a normal Claude chat, since the agent's summary box has no dictate option, and ask for the instructions only. We told it not to build anything, just to give us the prompt.

We said the agent should scan the projects board every week and listed every column to report on: project name, owner, the subitem status roll-up, project status, expected completion date, priority, revenue, cost, gross profit and date created. The report is for the management team, as a PDF emailed to a demo address at 8am every Monday.

We wanted a breakdown of how the projects are going and any risks, such as gross profit, a project getting close to its expected completion date or overdue subitem tasks, and left the rest to Claude. We also asked it to prepare a skill for the report's design.

Should You Give Claude An Example Report?

Yes, if you have one. Attach it, say it's the report the agent should produce, and ask Claude for a zip file containing a skill you can upload to the agent.

We had no existing report, so Claude designed one. It came back with the instruction set and a zip file for the skill, which includes an example report. Our board only had three projects, and the example was filled with dummy data.

The prompt took a long time to finish, and it seemed to be testing itself. We're not sure why.

Which Model & Tools Does A Weekly Report Agent Need?

We kept Opus 5.5 and connected two tools: monday.com to read the projects board and Gmail to send the report. A less capable model is cheaper but may write a worse report. A medium or low effort level is plenty for a job that runs once a week, which we expected to cost about a dollar a run.

The system prompt comes from Claude's instructions: once they're in the summary box, Generate writes the name, description and system prompt in far fewer words. That's deliberate, because the more detail an agent has, the longer it thinks and the more chances it has to get confused.

Both tools are on Always allow. Always ask would have the agent asking before everything it does, and Auto, which is new, has been unpredictable for us. Always allow lets it read and send email and build boards, which was low risk on our demo Gmail account, but on a live mailbox you may want tighter permissions.

How Do You Control What The Report Looks Like?

With a skill: packaged instructions and scripts the agent loads when it needs them. Our system prompt points the agent to it for the report's design, and it holds an example report plus instructions for building one. Here's how we added it:

1. Download the skill's zip file from the Claude chat.
2. Press Add skill, then Create skill, then Choose files, and open the zip file.
3. Create the skill. It appears alongside the predefined skills from Anthropic.
4. Go back to the agent and add the new skill to it.

Does The Report Agent Need An Advisor Or Sub-Agents?

We turned on an advisor and added no sub-agents. An advisor is a second model the agent turns to when it's confused, so use a more capable one than the model running the agent, much as you'd ask a manager who knows that particular job better.

The session logs show when the advisor was asked, but not what was said. Sub-agents are separate agents it can call on for a different job, and workflows (which plan the work and run agents in parallel) had only just been released, so we left them alone.

How Do You Set The Agent To Run Every Monday?

With a deployment on a weekly schedule, such as Monday morning. We set the trigger to manual while testing, so we could run it on demand, and the schedule goes on once it works.

Once you press Create agent, the agent needs an environment to live in. Claude's cloud is the easier choice unless you run your own servers. Unrestricted networking is simplest, since this agent isn't doing anything risky, but you could allow only package managers and MCP servers and list the allowed hosts.

The deployment takes the environment, the vault with your tool logins and an initial message sent at the start of every run (ours was "please run the report"). A budget is optional, but it ends the session when reached, finished or not.

How Do You Run The Agent & Watch It Work?

Press Run on the deployment to start a new session. Filter the sessions by that deployment and click in to watch the agent read the board and work through its instructions.

A successful run ends with an email carrying the PDF, built from the projects board and laid out the way the skill describes.

Can A Claude Agent Email A PDF Attachment?

Not in our test. The run cost $1.60 and built the report, but no email went out, so we told the agent in the session that it had a Gmail tool and asked it to send the attachment.

It asked its advisor for help, and the cost in the session jumped noticeably. It then reported that the Gmail tool can't attach a file from disk, only base64 text written into the call, which for this PDF was about 180,000 characters.

That's what testing is for. The instructions need to tell the agent it has the Gmail tool, and the file needs another route, such as saving it to Google Drive and emailing a link.

What Does The Finished Report Look Like?

It follows the skill's design, with the projects board's information laid out as we asked. The PDF was saved in the session log, so we could still download it even though the email failed.

From there it could go to the team, but it still needs reviewing, testing and refining until the instructions are exactly right.