What is Claude Code? The $200 Agent That Replaced My

By Lazar Milicevic · Published June 20, 2026 · 9 min read

If you're a founder paying €1,500/month to a contractor for ticket-level work — migrations, webhook fixes, PDF template edits — you're about to feel something uncomfortable. I shipped a VAT compliance refactor for my invoicing SaaS over one weekend, for $200 total, and I wasn't at the keyboard for most of it. Here's exactly how that session worked and why it matters for anyone running a small business.

Claude.ai vs Claude Code: one difference, entire story

Claude.ai is the chat in your browser. You paste a problem, it answers, you copy code back into your editor, you run it, it breaks, you paste the error back, you copy the fix back. You are the typist. You are the file system. You are the shell.

Claude Code is the same model — but it lives in your terminal with full filesystem access and a shell of its own.

That one difference is the whole story:

Browser Claude advises. Terminal Claude acts. Once you internalize that, everything about how you brief it changes.

Installation is one command:

npm install -g @anthropic-ai/claude-code
cd ~/projects/my-saas
claude

You're now in a session that sees your files, your test suite, your git history, your .env. Treat it accordingly.

The VAT exemption ticket: a real production session

I run Fakturko, an invoicing SaaS for small businesses in Serbia. Real paying customers. Real compliance rules. The ticket I gave Claude Code last weekend:

Add VAT exemption for EU B2B invoices. Validate VIES VAT IDs on customer creation. Generate the correct PDF footer per Article 196 of the EU VAT Directive. Update the Stripe webhook so exempt invoices don't double-charge tax. Write tests.

That is not a one-liner. That is:

• A schema change on the customers and invoices tables
• A database migration
• A VIES validation service
• Tax service logic changes
• PDF template edits (Jinja + WeasyPrint)
• Stripe webhook updates
• Unit + integration tests

Contractor budget for that scope, at the rates I was paying: roughly €1,200–€1,500 and 4–5 business days.

I opened a Claude Code session and pasted the ticket in plain English. First move it made — and this is the move most people skip — it read the codebase. Not just the file I pointed at. It pulled the invoice model, the tax service, the PDF generator, and the Stripe handler into its working context. Then it proposed a plan:

Plan:
1. Add `vat_exempt`, `vat_id`, `country_code` to customers table (migration)
2. Add `exemption_reason` to invoices table (migration)
3. Create services/vies_validator.py — call EU VIES SOAP endpoint
4. Update services/tax_calculator.py — branch on B2B + EU + valid VAT ID
5. Update templates/invoice.html — add Article 196 footer block
6. Update webhooks/stripe.py — skip tax_behavior=automatic on exempt
7. Add tests: tests/test_tax_calculator.py, tests/test_vies.py
Proceed?

I approved. Then I let it run.

What actually happened over 14 hours

• Wrote migration, ran it locally, failed on a foreign key constraint I'd forgotten about. Read the error output, rewrote the migration, ran it again, passed.
• Edited the tax service, added exemption logic, wrote unit tests, ran them — 2 failed. Read the assertion output, fixed the branching logic, re-ran, green.
• Edited the PDF template, regenerated a sample invoice, opened the PDF, noticed the footer was misaligned because of a missing CSS class, fixed the CSS, regenerated, correct.
• Updated the Stripe webhook, wrote integration test using stripe.WebhookEndpoint mock, passed.
• Committed each phase with a clean message.

I checked in between coffees. Approved the next phase. Went back to my life.

The economics: $200 vs €1,500

Here are the real numbers from that month, no estimates:

That $200 plan also covered three other shipping streams that month — a client ops automation, a lead-gen tool, and a build for a client. Cost per ticket is effectively rounding error.

The thing to understand: I'm not pricing this against "writing the code myself." I would have written it myself eventually. I'm pricing it against the contractor I'd otherwise pay for ticket-level work I don't want to be the bottleneck on.

Where it loses, and how you protect yourself

Claude Code is not magic. I've watched it ship beautifully and I've watched it confidently produce garbage. It loses on three specific things, and you have to handle each one:

Three failure modes you must manage

• Ambiguous specs. It will guess, and it will guess wrong. The fix is to write tickets like a product manager, not a tweet. "Fix tax bug" gets you garbage. The VAT ticket above worked because every constraint was named: which directive, which webhook, which behavior on exempt invoices.
• Design decisions. It picks the first reasonable pattern, not the best pattern for your codebase. The fix is a CLAUDE.md file at the project root with your conventions — folder structure, naming, test framework, ORM patterns, error handling style. Without it, you get a Frankenstein codebase across sessions.
• Security-sensitive code. You do not ship payment logic, auth, or anything touching PII without reading the diff line by line. The agent drafts. You approve. That's the contract.

A minimal CLAUDE.md that has saved me hours:

# Project conventions

- Python 3.11, FastAPI, SQLAlchemy 2.0 async
- Tests: pytest, async, in tests/ mirroring src/ structure
- Migrations: alembic, one file per logical change
- No print statements — use structlog
- All money values in cents (int), never float
- Stripe API version pinned in stripe_client.py — do not bump
- Run `make test` before claiming done

That file gets loaded into every session. It removes 80% of the "why did it pick that pattern" moments.

Stop using it like Copilot

This is the mindset shift that separates people who get $200 of value from people who get $5,000 of value out of the same plan.

Copilot completes your line. Claude Code completes your ticket.

If you're asking it to suggest a function name, you are paying $200/month for a $20 tool. The level it actually operates at:

• "Add OAuth2 login with Google. Update the user model, add the migration, wire up the callback route, write the session middleware, add tests."
• "There's a bug where invoices over €10,000 show the wrong total in the PDF. Find it, fix it, write a regression test."
• "Refactor the notification service to use a queue. Move the existing sync calls behind a worker. Don't break the existing API."

Each of those is a ticket. Each closes in one session. Each used to be a contractor day.

How I actually run it day to day

My setup, in case it's useful as a template:

• Ubuntu 22.04 on a home server (the "PC-PC" box), accessed over SSH from wherever
• Three to four parallel Claude Code sessions, each in its own tmux window, each scoped to one project folder
• One session shipping invoicing features on Fakturko
• One running a client ops automation (UNA_Intel — Gmail-Telegram bot ecosystem)
• One on the lead-gen tool (bizflowai.io-Catalyst)
• One on whatever client build is active that week

# typical morning
tmux attach -t shipping
# window 0: cd ~/projects/fakturko && claude
# window 1: cd ~/projects/una-intel && claude
# window 2: cd ~/projects/bizflowai.io-catalyst && claude
# window 3: cd ~/projects/client-XYZ && claude

One operator. Four shipping streams. Because the agent does the typing and I do the thinking. The bottleneck shifts from "how fast can I type" to "how well can I scope and review." That's the unlock.

Why bizflowai.io helps with this

If you're not a developer, running four parallel Claude Code sessions is not where you start — you start by identifying which repetitive ticket-level work in your business is even worth automating in the first place. At bizflowai.io I help small business owners install this layer: the agent-driven automations for invoicing, lead follow-up, client ops, and webhook plumbing that would otherwise cost €1,500/month in contractor time. We do the scoping, the conventions file, the safety review, and ship the working system into your stack.

The conversation that changes after this

If you run a small business and you're paying a contractor for ticket-level work right now, this is the layer to install before you hire your next person. Not instead of humans — before. Because once you see what one operator with an agent can ship in a weekend, the hiring conversation changes completely. You stop hiring people to type. You start hiring people to think, scope, and review. That's a different role at a different price point with different leverage.

The $200 isn't the story. The story is what that $200 lets you ship without becoming the bottleneck.

Frequently asked questions

What is the difference between Claude.ai and Claude Code?

Claude.ai is the browser chat where you paste text and copy answers back. Claude Code runs the same model in your terminal with full filesystem access and a shell. The browser version advises; the terminal version acts. It opens files, edits them, runs your build, reads errors, fixes them, and commits the result. You become the product manager instead of the typist.

How do I use Claude Code effectively on a real feature ticket?

Open your repo, start a Claude Code session, and describe the ticket in plain English like a product manager would. Let it read the codebase and propose a plan before touching anything. Approve the plan, then let it write migrations, edit code, run tests, fix failures, and commit. Check in between phases to approve next steps rather than typing alongside it.

Why does writing detailed tickets matter for Claude Code?

Claude Code fails on ambiguous specs because it will guess and guess wrong. Writing tickets like a product manager, not a tweet, prevents this. It also picks the first reasonable pattern rather than the best one for your codebase, so you must set conventions up front in a project file. For security-sensitive code like payments, always read the diff before shipping.

When should I use Claude Code vs GitHub Copilot?

Use Copilot when you want line-level autocomplete while you type. Use Claude Code when you want to close entire tickets, ship features, or fix bugs across multiple files with test runs. Copilot completes your line; Claude Code completes your ticket. If you're only asking Claude Code to suggest function names, you're paying for an agent and using it as a completion tool.

How much does Claude Code cost compared to hiring a contractor?

The Claude Max plan referenced in this workflow runs roughly two hundred dollars per month. In one example, fourteen hours of agent time over a weekend shipped a VAT exemption feature involving schema changes, migrations, PDF templates, Stripe webhook logic, and tests. The equivalent contractor work was estimated at around fifteen hundred euros, making the cost difference the core economic argument.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.