Using ChatGPT desktop or Claude desktop

Connect ChatGPT desktop or Claude desktop to your case, send the first prompt, and ask questions that work well across thousands of files.

#Before you begin

You need:

1. A case that has at least some files read (check Progress).
2. The command-line tool installed, from the Start Up screen.
3. The guide files written into the agent work folder (Write guide files on Start Up). Recommended.
4. ChatGPT desktop (in a Work session) or Claude desktop, set up so the agent can run commands on this computer. That's how it reaches the case: by running discovery.
5. An AI plan your firm has approved for client material. In this version the agent reads the case as written, with real names (see Stand-ins and what the AI sees).

Other agents that can run shell commands on the same computer, such as Codex or Claude Code, work the same way: open the work folder and paste the same prompt.

#Connect, step by step

1. In Discovery Engine, open Start Up and click Copy prompt.
2. Open ChatGPT desktop and start a Work session, or open Claude desktop.
3. Choose the agent work folder as the project (Start Up shows its path; by default Documents/Discovery Engine/Agent work).
4. Paste the prompt as your first message and send it.
5. If the agent app asks for permission to run commands, allow it for the discovery command.

The prompt asks the agent to read the full guide (discovery agent-guide), run discovery doctor and discovery status, and then tell you in a few lines how many files are searchable, whether anything is still being read, failed or blocked, and what it can do for you. Then it waits for your questions.

#Asking questions

Ask the way you'd ask a capable paralegal who has instant access to every file. The agent searches, reads the passages that matter, and answers with citations.

Good first questions:

• "When did Daniel Reeves and Marisol Vance first communicate? Cite everything."
• "Which documents contradict the statement that the wire was sent on March 3, 2019?"
• "List every payment to Harrowfield Holding Co. in 2019, with the document and page for each."
• "What does the recorded call on March 2 say about the account ending 4471? Give me the timestamps."
• "Summarize the e-mail thread about the quarterly figures, oldest message first."

#What a good answer looks like

• Each claim names the file and the citation: wire-request-form-2019-03-03.pdf, page 1 (DOC-0095 p.1).
• It ends with a Sources section: one line per document, with the file name linked to the original, the citation linked to the document in the app, and the file's path.
• If it couldn't find something, it says how it searched and how much of the case the searches covered, rather than guessing.

Click a citation link to check it in the app. If the chat window won't open discovery-engine:// or file:// links, use the path printed beside them, or find the document by ID on the Documents screen.

#Questions across many documents

A search returns the best passages, not every one. For questions like "every payment to…" or "who knew about the loan, and when", ask the agent to be thorough, and it has tools for it:

• Counting first. It can count the documents that hold certain words by month, year, kind or person, and list every one of them.
• Overviews. It can look at many documents at once as short "cards", list the people in the case and who appears with whom, put documents on a timeline, and show a whole e-mail thread in order.
• Reviews. When the answer must rest on every document that could hold it, it starts a review: a queue of documents in date order. It reads each one and records a call (relevant, not relevant or unsure) with a reason and citations. Reviews are kept in the case, so a later session can pick one up.
• Connections. It can show everything linking two people (passages naming both, e-mails between them, shared accounts or addresses, chains through others) and work through a connections study you started on the Connections screen.

Try:

• "Start a review of every document mentioning the loan before May 2019 and work through all of it. Tell me how many you reviewed."
• "How are Daniel Reeves and Priya Raghunathan connected? Look past documents that merely mention both."
• "Work through review 3 and tell me when you're done."

#Highlights for the attorney

Ask the agent to propose highlights:

• "Read the issue list, then propose highlights for each issue. Only highlight passages you've read."

It proposes each passage with an exact quote, a citation and a one-line reason. The app checks every quote against the document. The proposals then wait on the Highlights screen for you: the agent cannot approve them. See Highlights.

#Exports

When you're ready, ask: "Export the highlighted PDFs and the exhibit index to the export folder." The agent asks before exporting (the prompt tells it to) and writes to the export folder from Start Up. You can also export yourself on the Exports screen, which is usually simpler. See Exports.

#While the case is still reading

You can start before reading finishes; the agent works on what's searchable so far. Ask "How is reading going?" and the agent checks the status and the activity log, then tells you how many files are done and left, the time left, and which file is being read now. If reading seems stuck, it looks for the log's heartbeat before saying so.

#When the agent says it needs you

#Starting a new chat

Each new chat starts fresh. Paste the prompt again (copy it fresh from Start Up, since case counts change). Anything the agent saved in the work folder (its notes) and in the case (reviews, notes, labels) is still there for it to pick up.