Skip to main content
Same portfolio question. Two doors. You will feel why web chat and Desktop Cowork are not interchangeable for BizOps work. The deal is fictional on purpose: Harborview Plaza, three retail/office renewals heading into IC. You and your facilitator run both paths together. Copy the prompts, paste them, and check the expected outputs before you move on. Plan for about 30 minutes. You need Claude on the web (claude.ai) and Claude Desktop signed in on the same account.

​
Download the Harborview pack first

Do this before either run. The zip is the lab kit: deal sheet, rent roll, and three lease abstracts. Unzip it to your Desktop so Cowork can see a real folder. Sample pack

Unzips to a folder named Harborview-lab with six training files. Fictional data only.

Download zip
  1. Click Download zip.
  2. Unzip it. On macOS, double-click the file. On Windows, right-click → Extract All.
  3. Move the resulting Harborview-lab folder to your Desktop (or leave it in Downloads if you prefer).
  4. Open deal-sheet.txt in any text editor and skim it for sixty seconds. You should know the three tenants: Sweetleaf Juice, Paper & Parcel, Northline Dental.
Part 1

Web chat + research for market context

10 min
Part 2

Desktop Cowork against the local pack

14 min
Part 3

Score which surface won, and why

6 min

​
The deal you are working

Northline Realty wants a Friday IC recommendation on three upcoming renewals. The question is not “can Claude write a memo.” The question is which surface gets you to a defensible recommendation with less friction.
Surface A · Web

Best when the missing ingredient is outside your laptop: market comps, public context, a fast memo from a pasted deal sheet.

Surface B · Desktop Cowork

Best when the evidence already lives in files: rent roll, lease abstracts, and a workbook or memo Claude should assemble without you shuttling pages.

​
Part 1: Web — research the market, draft the memo

Do this in the browser at claude.ai/new. Stay in Chat, not Cowork.
  1. From your unzipped Harborview-lab folder, open deal-sheet.txt and keep it handy.
  2. Start a new chat.
  3. Turn Research / web tools on.
  4. Attach deal-sheet.txt (or paste its contents), then send Prompt A.

Web Prompt A — market context for Harborview

​
Expected output — Web

You are looking for shape, not perfection. Your run should clear most of this list:
Citations

At least two real sources, dated or clearly current, about retail rent or renewal conditions.

Per-tenant posture

Sweetleaf, Paper & Parcel, and Dental each get take / counter / walk plus a rent floor.

Concession logic

One trade and one refusal per tenant, tied to the deal sheet (TI, free rent, exclusives, dock access).

IC memo

About 180 words, decision-first, no hype, readable aloud in under two minutes.

Strong runs usually counter Sweetleaf (juice sales support rent, but free rent is expensive), treat Paper & Parcel as the softest credit (late payment + TI ask + no option), and take or lightly counter Dental while flagging the exclusive. Weak runs invent exact comps with fake precision, or write a long essay with no posture table.
Leave this chat open. You will compare it to Cowork in Part 3.

​
Part 2: Desktop Cowork — work the file pack

Now switch machines, not accounts. Open Claude Desktop, same login.
  1. Confirm Harborview-lab is on your machine from the zip download above. If you skipped it, download the pack and unzip it now.
  2. Start a new conversation. Select Cowork in the composer.
  3. Connect / point Cowork at the Harborview-lab folder (not the zip file).
  4. Send Prompt B. Do not paste the files by hand. The point is the folder.

Cowork Prompt B — build the renewal pack from the folder

Watch three moments while it runs:
  1. Plan — Claude proposes steps before writing files. Edit the plan if it tries to invent market comps instead of using the folder.
  2. Approvals — when it asks to create files, read the path and action, then allow.
  3. Deliverables — open the three files on disk. If they only exist as chat text, the run failed the Cowork test.

​
Expected output — Cowork

Real files

Matrix, IC brief, and open-questions land in the Harborview folder, not only in the chat transcript.

Matrix rows

Three renewal suites, with posture and rent floor filled from the pack.

Open questions

At least three human checks (for example: verify Sweetleaf sales, confirm vacant bay asking rent, legal review of dental exclusive).

No silent deletes

Source abstracts and rent roll remain untouched.

Strong runs create files on disk, keep judgments tied to suite-level evidence, and put unverified items in open-questions.md instead of hallucinating answers. Weak runs paste a beautiful memo into chat and never write the folder, or “helpfully” rewrite the source abstracts.
Optional stretch if you finish early: stay in Cowork and ask it to add a fourth file, landlord-counter-email.md, with three short tenant-ready counter emails. Still no sending.

​
Part 3: Score the surfaces

You now have two artifacts from the same deal. Mark which surface won each job. There is no perfect score. The point is to feel the grain of each door.

​
Routing rule to keep

Web wins when the bottleneck is outside context: market research, a pasted brief, a memo you will edit in chat.

Desktop Cowork wins when the bottleneck is file work: many local documents, a spreadsheet to build, a pack that should exist on disk when you walk into IC.

​
Flight card

​
Debrief with your facilitator

  1. Where did web feel faster than Cowork, and where did it feel thinner?
  2. Did Cowork’s value come from the model, or from not having to shuttle files?
  3. Look at next week’s real work. Name one web job and one Cowork job. Be specific about the folder or the research need.
  4. What would have to be true before you pointed Cowork at a live deal room?