How do I summarize a set of project documents with an agent?
Place source documents in a local workspace, name the files and output structure, request source references, and check the summary against the originals.
Maintained by the 路墨 Lumoswitch product team
Open a local folder containing the materials, specify inputs and output format, and keep the originals for verification.
How should I prepare the documents?
1. Make a local folder for this set of materials, copy the files you want summarized into it, and select that folder in the workspace. 2. Start with a few readable text documents. If a format cannot be read, export readable text from its original app; do not assume every scan or attachment can be parsed. 3. Keep the originals unchanged and avoid placing unrelated private material in the workspace.
What task can I give the agent?
Name the files, audience, and output format. For example: “Read only `meeting-notes.md`, `requirements.md`, and `open-questions.md`. Prepare a one-page summary for the product team: start with conclusions, then list confirmed decisions, open questions, and owners/dates (write ‘not specified’ when the source does not say). Cite the source file for each point. Do not add facts that are not in the files, and do not edit the three inputs. Write the result to `summary.md`.” 1. Use Plan mode when you need to compare several sources and track references. 2. Include the exact files, summary structure, and output path in the completion criteria. 3. If the app asks to create the output file, verify the path before approving.
How can I check that the summary is accurate?
1. Open the output and spot-check each conclusion against the cited file and passage. 2. Verify dates, numbers, owners, exceptions, and anything described as undecided against the original. 3. Confirm that the inputs were not changed; the run's file-change record should normally contain only the summary you requested. An AI document summary is not a legal, financial, or contract review.
What if the agent cannot find a file or returns very little?
1. Confirm that the selected folder contains the materials, rather than its parent or a different project. 2. Check that the filenames in the task match the actual names and capitalization. 3. For scanned PDFs, images, or complex formats, export selectable text from the original app or provide the relevant passages in smaller batches. 4. If the request is too large, summarize one or two files at a time and ask the agent to state what it could not read. Do not ask it to guess.