AI for Designers

Twelve AI Workflows for Designers

Guide 02: Use Claude's Computer for Real Files, Real Code, Real Deliverables

Claude can operate an actual Linux machine: clone repositories, install packages, run analysis, and hand you finished Word docs, spreadsheets, slide decks, and PDFs instead of text you have to reassemble yourself.

Difficulty: Starter · Time: 10 min setup, useful immediately You'll need: Claude with the code execution / file creation feature enabled (Settings → Features on claude.ai; availability varies by plan)

What changes

When this feature is on, Claude's answers can end in artifacts you download rather than text you copy. Behind the scenes Claude gets a sandboxed computer where it can write and run code, install packages, and produce files. For designers, the everyday wins:

The mental shift: stop asking Claude to tell you things and start asking it to make things.

What you need

Toggle on the file creation / code execution capability in Settings → Features. Check current plan availability at https://support.claude.com if you don't see it. That's it; the computer, the packages, and the document tooling are all Claude's problem.

One thing worth knowing about the sandbox: it resets between tasks, and its network access is limited to safe destinations like package registries and GitHub. Your files exist there only for the duration of the work.

First run: three starter workflows

A. Survey data → analysis deliverable

  1. Export your survey results as CSV and attach the file to a chat.
  2. Prompt: "Analyze this survey export. Clean obvious junk responses, then give me: response distributions per question, any statistically meaningful differences between user segments, and three charts worth showing stakeholders. Deliver as an Excel workbook (data + pivot summaries) plus a one-page PDF readout."
  3. Interrogate the analysis before trusting it: "How did you decide what counted as a junk response? Show me the rows you excluded."

B. Outline → formatted document

  1. Paste your rough notes or outline.
  2. Prompt: "Turn this into a formatted Word document: title page, headings, a summary table of findings, page numbers. Professional but not sterile."
  3. Iterate on the file like any draft: "Move the recommendations before the methodology, and tighten the executive summary to 150 words." Claude edits the actual file and returns a fresh copy.

C. Repo → understanding

  1. Prompt: "Clone [public repo URL] and explain how its component library is organized: folder structure, how theming works, and where design tokens live. Write it as an onboarding doc for a designer who reads a little code."
  2. This is the quiet superpower for designers working with engineering: you can ask questions of the codebase itself instead of waiting for someone to explain it.

Example prompts

Attached are 40 usability session note files. Merge them, extract
every observation tagged "issue", deduplicate similar ones, and
give me an Excel sheet: issue, frequency, severity, sessions
where observed, representative quote.

Write a Python script that takes a folder of images and exports
each at 1x/2x/3x with a consistent naming scheme, then run it on
the images I've attached and give me the outputs zipped.

Build a simple interactive prototype of a drag-to-reorder list
with the physics I describe below, as a single HTML file I can
open locally and put in front of users: [description]

Troubleshooting

"I don't see the file." Ask Claude to present the file explicitly. If a task ended with only a description of a file, the last step got skipped; a nudge fixes it.

The document formatting is fine but bland. Direct it like you'd direct a designer: reference a structure you like, specify typography weight and restraint, ask for less bolding. Default document styling is conservative on purpose.

Analysis results you can't verify. Always ask for the working: which rows were used, what test was run, what assumptions were made. Claude will show the code and intermediate data; reviewing it is the QA step.

A package or website can't be reached. The sandbox's network is restricted. Provide the material as an attachment instead, or ask Claude which alternative it can use.

Where this takes you

This capability is the engine room for more advanced workflows: accessibility audit reports, design-token transforms, and design system changelogs all become downloadable deliverables through it. The habit that compounds: end substantial working sessions with "package this up" (analysis, decisions, and artifacts in one file) and you never lose a session's output to chat scroll again.