Explore a dataset and write up findings

Uses

• What the agent can see
• Workspace agent
• Data files
• File workspace

By the end of this guide you will have a project with a CSV file, an AI conversation that read it, and a short markdown writeup whose conclusion is grounded in the agent’s answer.

It walks the smallest credible version of exploratory data analysis with narration: data in the project, a grounded agent answer, and prose that cites that work.

The critical detail most users miss: the agent does not automatically read the file you have open. It sees the project’s file tree and the path of your active file, but not the file’s contents. To make the agent read the data, you have to attach the file to the chat message. This guide makes that step explicit.

Prerequisites

• About ten minutes.
• Any modern browser.
• A CSV file. If you do not have one, copy the snippet below.

Worked example

The CSV is small on purpose so you can verify the agent’s answer by eye.

date,product,units_sold
2026-01-04,bagel,42
2026-01-04,muffin,18
2026-01-05,bagel,55
2026-01-05,muffin,12
2026-01-06,bagel,31
2026-01-06,muffin,27

Save it as sales.csv on your computer, or you will drop the snippet directly into the project in step 2.

[Asset: short GIF] One screen recording, around 30 seconds, showing the attach → ask → answer loop with the inline CSV. Asset name: data-loop-attach.gif. Replaces no existing placeholders.

Steps

1. Create the project

1. From the dashboard, click New project.
2. Pick the Markdown preset and click Let’s start.

A new workspace opens. The header storage icon shows a browser glyph, which means the project is private to this device until you share it.

2. Add the CSV

Either drag sales.csv from your computer into the file tree, or create a new file called sales.csv and paste the inline snippet above.

Click sales.csv in the file tree to open it. Titan opens CSV files in the data viewer rather than a code editor. You should see six rows in a table.

3. Open the chat and attach the CSV before asking

This is the step most users skip. Skipping it means the agent answers from the file name alone and may make up content.

1. Open the right sidebar and click AI Chat.
2. In the chat composer, click the attachment button.
3. Pick sales.csv from the file picker.

You should see an attachment chip appear above the prompt input, labelled sales.csv. Until the chip is there, the file’s contents are not in the agent’s context.

4. Ask a question grounded in the data

With the attachment chip visible, type a question that requires reading the rows. For example:

Which product sold more units across the three days, and by how much?

Send.

The agent should return an answer that names both products and gives concrete numbers (in this dataset, bagels sold 128 units against muffins’ 57, a difference of 71). If the answer is vague or invents numbers that are not in the data, the file was not attached — check the chip.

5. Turn the answer into a written paragraph

1. In the file tree, click + and create a new file called findings.md.
2. Open it. The Markdown preset opens new .md files in Visual mode. You can switch to Code from the status bar if you prefer raw markdown.
3. Add a heading and write a paragraph that uses one of the numbers the agent gave you. You can quote the agent or paraphrase.

Save. The project now contains the data, the conversation, and the writeup.

Verification

You have completed the loop if all of these are true:

• The agent’s answer in step 4 contains specific numbers that match the data (you can spot-check by eye against the six rows).
• The attachment chip for sales.csv is visible above the prompt input.
• findings.md contains a paragraph that references at least one number from the agent’s answer.
• Reopening the project after a refresh shows all three files (sales.csv, findings.md, your conversation in the AI panel).

Variations

• Use your own larger CSV. The flow is identical. The agent reads the whole file you attach (within attachment-size limits).
• Use multiple files. Attach two or three files in the same message. The agent reads all attachments.
• Have the agent draft the markdown for you. Ask “write a one-paragraph summary suitable for a notes file” instead of asking the analytic question, then accept the proposed change. See Agent actions and approvals for how the changeset review works.
• Pin the conversation per dataset. Each project keeps its own conversations. If you have one dataset per project, the conversation history stays scoped to that data.

If something goes wrong

• The agent’s answer is vague or invents data. The file is not attached. Check for the attachment chip above the prompt input before sending.
• The CSV opens in a code editor, not a table. The extension may be wrong. Files routed to the data viewer must end in .csv (or another supported data extension; see Data files).
• The attachment button is greyed out. No model is enabled. Open the model selector and pick or enable a model. See Models and AI settings.
• The agent’s answer is correct but cannot edit findings.md. The selected model is in chat-only mode. To have the agent propose file edits, switch to a model with tool support.
• You expect the agent to read more than what is attached. It will not. The agent’s context is exactly the file tree, your active file path, prior messages, and the files you attached. See What the agent can see.

Read next

• What the agent can see — the canonical scope of agent context.
• Workspace agent — the full reference for the AI panel.
• Data files — formats, viewer features, and limits.