Beyond Code

What Anthropic Discovered

I’m writing a series breaking down how Claude Code works—how it’s built, how people actually use it, and what it means for the future of software. This article stands on its own, but you can also start from the beginning or catch up with the previous entry.

Previous: The AI That Writes Itself

Anthropic built Claude Cowork in 10 days. They did this, because their users taught them something they didn’t expect.

Inside Anthropic, engineers started noticing something strange. Data scientists with Claude Code running on their machines. “I figured out how to get this thing running and write queries for me,” one explained. Writers were organizing research. Admins were processing documents. Researchers were summarizing papers.

Nobody told them to do this. No tutorials suggested it. They just discovered that a tool with filesystem access and shell commands is useful for any work that involves files.

The same pattern emerged independently among external users. Writers, researchers, analysts — all repurposing a “coding tool” for non-code work.

Convergent behavior. The kind of signal that screams “there’s a product here.”

So Anthropic built one. 10 days, idea to launch. That’s how clear the signal was (and how well their product development pipeline functions).

But here’s the thing most people miss: Cowork exists because Claude Code’s capabilities were never really about code. The name is misleading. The architecture tells a different story.

Why This Matters

Many people dismiss Claude Code when it comes to their work because of the name. They assume it’s for developers, move on, and miss what might be the most capable AI assistant available today.

If you work with files — any files — Claude Code can help. Documents, data, images, notes. If it lives on your computer, Claude can read it, process it, and write the results somewhere useful.

This is knowledge work — any kind.

Why It Works for Non-Code

Claude Code’s architecture doesn’t care what kind of files you work with:

• Filesystem access: Read and write any file format

• Shell execution: Run any command-line tool

• MCP connections: Access external tools and data

• Skills: Define any workflow, not just coding workflows

The same capabilities that make it powerful for coding make it powerful for everything else.

Content Work

Writing, research, and documentation all follow the same pattern: read source materials, process them, produce output.

Ask Claude to “write a blog post based on these research notes” and it reads your notes folder, synthesizes the content, and saves a draft. Ask it to “compare these three articles and highlight differences” and it does exactly that — no copying and pasting between windows, no manual summarization.

The prompts are natural language. The results are files you can use.

Data Work

You don’t need to be a data scientist to work with data. You just need to describe what you want.

“Parse this CSV and show me trends” works. So does “convert this Excel file to clean CSV” or “merge these datasets and remove duplicates.” Claude figures out the commands, runs them, and gives you results.

For anything more visual — charts, graphs, reports — Claude can generate the code or connect to tools that create them. You describe the output; it handles the process.

File and System Work

The mundane stuff is where Claude Code quietly shines.

Organizing photos by date. Renaming files to match a convention. Finding duplicates. Converting formats. Cleaning up old directories. These tasks are tedious for humans and trivial for an AI with filesystem access.

Same for system administration: checking disk usage, analyzing logs, writing backup scripts. You could learn the commands yourself, or you could describe what you want and let Claude handle it.

Automating Your Workflows

Skills take this further — reusable workflows you define once and invoke whenever appropriate.

# .claude/skills/daily-summary/SKILL.md
---
name: daily-summary
description: Create a daily summary from notes
---

1. Read all files in ~/notes/daily/
2. Extract key decisions and action items
3. Write summary to ~/notes/summaries/YYYY-MM-DD.md

Type /daily-summary and it runs. Your workflow, captured and repeatable.

Connecting to External Tools

MCP lets Claude Code talk to services beyond your filesystem:

• Notion for pages and databases

• Google Drive for documents

• Slack for messages and reports

• Calendar for scheduling

“Check my Notion for open tasks and create a summary” becomes a single request instead of a multi-step manual process.

The Real Limit

Claude Code’s power comes from a simple stack:

1. File access — Read and write anything

2. Shell access — Run any tool

3. MCP — Connect to external services

4. Skills — Encode any workflow

5. Natural language — No programming required

The limit isn’t what Claude Code can do. It’s what you think to ask.

If you work with files and haven’t tried it — you’re leaving capability on the table.

---

What’s Coming

Next up: The Numbers — Claude Code’s growth, revenue, and what the stats tell us about where AI tooling is headed.

If that sounds useful, subscribe. I’ll see you in the next one.