Skip to content

Demo 2 · Cowork

Abstracts to spreadsheet

In 10 minutesFifteen paper abstracts — one slice of the chain-of-thought literature — turned into a screening-ready literature matrix in Excel, by one instruction.

No Cowork tab? See the checks on the start page. And the bottom-left should say University of Chicago Enterprise — yes, at TTIC; if not, do Sign in once first.

Download the abstracts (ZIP, 14 KB)

Steps

  1. Download the ZIP, then in your Downloads folder right-click → Extract All… → Extract. You'll get an abstracts folder with 15 text files, right next to the ZIP. (Double-clicking the ZIP looks like a folder of files, but it isn't one — Cowork can't use it. Extract first.)
  2. In Claude Desktop, switch to the Cowork tab and start a new session.
  3. When Cowork asks which folder to work in, choose Downloads → abstracts.

  4. Paste this and hit Enter:

    This folder contains 15 text files, each the title, authors, year, and
abstract of one research paper on chain-of-thought reasoning. Read every
file and build an Excel file called LitMatrix.xlsx with one row per paper:
File, Title, Year, Problem, Approach, Benchmarks, Key claim. Keep Approach
to five words or fewer, and under Benchmarks list only datasets the
abstract actually names. If an abstract doesn't state something, put NOT
STATED in that cell instead of guessing. Add a second sheet grouping the
papers into 3-5 approach families, with one line on what each family does.

    Now watch. Cowork opens each file, reads it, and builds the matrix. Along the way it may pause to ask a question or for a go-ahead — read what it says and answer in plain English; that is the skill. If it seems to stall partway, type continue.

    Cowork working through the abstracts

    Cowork lays out a plan and works the checklist — read the files, extract the fields, group the families, build the file, verify it — pausing to run each step.

  5. When it finishes, open LitMatrix.xlsx from the folder.

    The finished literature matrix

    LitMatrix.xlsx — one row per paper (Title, Year, and the fields you asked for), plus a second sheet of approach families. In this run, four abstracts named no benchmark — scroll right in the real file and those Benchmarks cells read NOT STATED.

No Excel on this machine? Go to office.com, sign in with your UChicago account, and upload LitMatrix.xlsx — it opens in Excel for the web.

Yours won't match this cell-for-cell — abstracts are advertisements, and a few of these genuinely don't state what they tested on. That's the point of the next step.

Don't take the first answer — push back

You'd never build a related-work section from abstracts you hadn't questioned. Same here:

Abstracts oversell. For each Key claim, quote the exact sentence it came
from, and flag every row where you interpreted rather than read. Then tell
me: which three of these papers would you read in full first, and why?

Claude shows its evidence, you make the call, it fixes the sheet. That back-and-forth is the whole job — you own the result, Claude does the typing.

Yes, this is allowed — it's the University of Chicago enterprise workspace. For unpublished data or anything under review, check UChicago's AI guidance first.

Your turn

Any folder of anything → any file you wish existed. The 40 PDFs in your to-read folder → a screening matrix like this one. A semester of meeting notes → an action-item tracker. The reviews on your last submission → a point-by-point response skeleton. Describe the end product, keep the "if it doesn't say, say NOT STATED" guardrail, and let Cowork do the walking — then check its work.

When the task is bigger — real data you'll re-run every month — that's an Automate. Last page.

Next: 3 · Code — kill a recurring data chore

Paper metadata (titles, authors, abstracts) from arXiv.org; each file cites its arXiv ID and link. The papers themselves remain under their authors' licenses.