Writing Cowork Prompts: Describe the Outcome, Not the Steps

Knowing what Cowork can do is one thing. Getting it to do what you actually want is another. The gap between a vague request and a useful result almost always comes down to how the prompt is written. This lesson covers the practical mechanics: what to include, what to leave out, and how to fix a result that is close but not quite right.

How Cowork prompting differs from chat prompting

In a standard chat conversation with an AI, you typically break work into steps yourself. You send one message, read the response, send a follow-up, and so on. You are the coordinator.

Claude Cowork changes that relationship. You describe the finished result and Cowork handles the steps. That is the core design. It means the prompt you write is less like a task list and more like a project brief. You are handing off the coordination, not just a single instruction.

This distinction matters in practice. A chat-style prompt (“First summarize the document, then extract the action items, then format them as a table”) can work, but it is doing Cowork’s job for it. A better prompt tells Cowork what the table should look like when it is done and lets it figure out the path.

Describe the outcome, not the steps

The most reliable way to write a Cowork prompt is to describe the finished result as precisely as you can. What does the output look like? What does it contain? What would make it wrong?

Vague: “Go through my invoices and organize them.”

Specific: “Review the invoice files in the Finance/2025 folder and produce a spreadsheet with four columns: date, vendor name, invoice number, and amount. One row per invoice. Sort by date, oldest first.”

The specific version tells Cowork what done looks like. The vague version leaves that judgment entirely up to Cowork, which means you are likely to get something that is technically correct but not what you needed.

Name the output format

Format matters. If you want a spreadsheet, say spreadsheet. If you want a markdown file, say markdown. If you want a summary paragraph, say that. When you name the format, also name the structure. Column headers for tables. File name and location for documents. Length for summaries.

Vague: “Summarize my meeting notes.”

Specific: “Write a 150-word summary of the meeting notes in Notes/June-15.txt. Plain prose, no bullet points. Save it as a new file in the same folder called June-15-summary.txt.”

You will save several rounds of back-and-forth by being specific about format up front.

Point at files rather than pasting them

Cowork works inside the folder you connect and can read files directly from there. You do not need to copy and paste file contents into your prompt. Pasting large blocks of text adds clutter and can push important parts of your instructions out of focus.

Instead, tell Cowork where the file lives. “The source data is in Reports/Q1-sales.csv” is better than pasting several hundred rows of spreadsheet data into the prompt box. If you have multiple files to work from, list them by path.

Scope tightly

Broad requests produce broad results. If you want Cowork to focus on one thing, tell it one thing. “Audit my entire website and improve everything” is a hard prompt to act on well. “Review the three product pages in /pages/products and rewrite the first paragraph of each so it leads with a specific benefit rather than a general description” is something Cowork can execute cleanly.

Tight scope also makes the plan easier to review. Cowork shows you a plan before taking significant actions, and a clear request produces a clear plan. If the scope is vague, the plan will be too, and you will spend more time approving steps you did not intend.

Two more before-and-after examples

Vague: “Write some social posts about my new product.”

Specific: “Write three LinkedIn posts about the launch of our Q3 inventory report feature. Each post should be under 150 words. Tone is direct and practical. No hashtags. Save them in a file called linkedin-drafts.txt in the Marketing folder.”

---

Vague: “Clean up the customer list.”

Specific: “The file Contacts/customers-raw.csv has duplicate rows and inconsistent capitalization in the Name column. Remove duplicates (keep the row with the more recent date), fix the capitalization to title case, and save the result as customers-clean.csv in the same folder.”

In both cases the specific version gives Cowork enough to work from without walking it through every step. The outcome is clear. The format is named. The files are identified.

Iterate by refining, not restarting

If the first result is not quite right, add a correction in the same session rather than starting fresh. Cowork retains context. A short follow-up (“The summary is good but too long, cut it to 100 words” or “Add a fifth column to the spreadsheet for the payment status”) is usually faster than rebuilding the prompt from scratch.

Restarting makes sense when the original request was fundamentally off. For most small corrections, refine in place.

For a deeper look at prompt structure that applies across AI tools, see the guide to writing AI prompts. If you want to test your prompts before running them, the free Cowork Prompt Builder lets you draft and check your requests in a structured format.

The next lesson covers permissions and safety settings, which control what Cowork can access and act on in your environment.