How to Use Claude Code for Business Automation (A Practitioner's Guide)

Most guides to Claude Code start with "install the CLI" and end with "here's how to write a prompt." That's not this guide.

This is about what comes after the first prompt works: how to build automation that runs reliably, recovers from failures, and produces results you can trust — without you watching it.

We run a company on this stack. Here's what we've learned.

What "automation" actually means with Claude Code

Claude Code is not a workflow builder like n8n or Zapier. It doesn't have a visual canvas, pre-built integrations, or a trigger system.

What it has: a capable AI agent that can read files, run shell commands, call APIs via curl or Python, write and modify code, and chain those actions together based on natural language instructions.

That means the automation you build with Claude Code is more flexible than a fixed workflow tool — but it requires more explicit specification. The agent doesn't have a pre-built "send Slack message" node. It has the ability to figure out how to send a Slack message if you tell it what you want to achieve.

This is a trade-off. More flexibility, more specification work.

The three patterns that work

Pattern 1: The task loop

The simplest reliable automation pattern: give the agent a task with a clear done condition, let it execute, verify the result.

TASK: Check the drafts directory for any .md files without a 'slug' field in frontmatter.
For each file missing 'slug': add a slug derived from the filename (lowercase, hyphens).
DONE WHEN: zero files in drafts/ lack a 'slug' field. Verify by grep before reporting done.

The key elements: what to do, what done looks like, how to verify. Without the verification step, the agent reports done when it finishes the last action — not when the goal is achieved.

Pattern 2: The check-and-act loop

For monitoring tasks: check a condition, act only if the condition is met, report what happened.

TASK: Check Stripe for new payments in the last 24 hours.
IF new payments exist: log them to payments.log with timestamp and amount.
IF no payments: log "checked at [time], no new payments".
DONE WHEN: payments.log updated with today's check.

The explicit "if no payments" branch is important. Without it, agents sometimes skip the logging step when there's nothing to log — and you can't tell if the check ran or was skipped.

Pattern 3: The pipeline

For multi-step processing: each step takes input from the previous, produces a file or state that the next step reads.

Step 1: Fetch new items from [source] → write to raw/new_items.json
Step 2: Read raw/new_items.json → filter by [criteria] → write to processed/filtered.json  
Step 3: Read processed/filtered.json → format as email → write to outbox/draft.txt
Step 4: Verify outbox/draft.txt exists and is non-empty → report path

Each step has a concrete output. If step 2 fails, step 3 has no input and fails loudly rather than silently. This is deliberate — you want failures to be loud.

What doesn't work (and why)

Ambiguous done conditions. "Clean up the repository" is not a task. "Remove all .bak files from src/ and verify with find" is a task.

No recovery path. If your automation hits an API rate limit, does it retry, wait, or crash? Specify. The agent will make a choice if you don't, and it may not be the choice you'd make.

Trusting the report over the state. An agent that says "done" has finished executing. Whether the goal was achieved is a separate question. Build verification into every task.

Sessions that run too long. Claude Code sessions have context limits. For automation that runs multiple hours, checkpoint state to disk at each step. Don't rely on the agent remembering step 1 by the time it reaches step 10.

When not to use an agent

Not every automation belongs in Claude Code. Prefer a simpler tool when:

• The task is purely mechanical: copy file, rename, move. A shell script is more reliable and cheaper.
• The task has a perfect existing integration: Zapier has 6,000 app connectors. If your automation is "when form submitted, send email," use Zapier.
• The task needs to run in under a second: LLM inference has latency. For real-time triggers, use webhooks and dedicated services.

Claude Code adds value when the task requires judgment: evaluating content, adapting to variable inputs, deciding between options based on context. That's where the cost of LLM inference pays off.

Getting started

1. Start with one task that has a clear, measurable done condition
2. Run it once, watch the output
3. Identify where it fails or produces unexpected results
4. Add verification and recovery handling
5. Only then automate it as a recurring job

The learning curve is front-loaded. Once you have a working pattern, extending it is fast.

Claude Code Mastery covers the patterns above in depth, with examples from our own production stack. Starting at €29.

Bratschke Solutions GmbH — agentic-movers.com | @spocky_magic_ai