I've been using Claude Code for months, and most people are missing the point entirely: it's not a tool for writing code faster. It's an automation platform. Paired with Playwright, webhooks, or its new "routines," Claude Code can drive a browser, qualify leads, respond to support tickets, all without your laptop being open. What I've come to understand is that the real question isn't "does AI code well?" but "have I built a system around it?"
- 🔑 Playwright + Claude Code automates anything that runs through a browser, even without an API.
- ⚠️ "Routines" are cloud agents: they run 24/7 without your computer being on.
- 💡 The real lever isn't the tool, it's the production system built around it.
- 🚀 A retainer-based agency model at $2,500/month per client becomes viable thanks to Claude Code.
Claude Code isn't just for generating code
When people talk about Claude Code, the conversation quickly turns to the quality of the generated code. Is it clean? Is it maintainable? That's a valid question, but it misses the main point.
The real power of Claude Code is its access to tools. It reads files, executes commands, queries APIs, controls a browser. This ability to act within a real environment is what sets it apart from a simple code generator.
A developer on Reddit put it simply: "don't confuse the code with the algorithm." Code is the mechanical part. What you decide to automate, that's where the freedom is unlimited. This distinction is at the heart of what I explore in the article on AI agent fundamentals.
Playwright + Claude Code: automating anything that runs through a browser
The most immediately powerful use case: connecting Claude Code to Playwright CLI to drive a browser. The principle is straightforward. You describe what you want done, Claude writes the script, executes it, observes the results, and fixes things if they break.
What's impressive is the learning cycle. With each iteration, the script becomes more precise. After four or five attempts, the agent knows where to click, handles pop-ups, and can even vote in a poll it had never seen before. Nate Herk documented this on YouTube with three concrete examples.
Practical applications range from automated QA testing on a multi-page app, to contact research (phone numbers, emails) on sites without an API, to authenticated sessions using your Chrome cookies. None of these scenarios require official data access.
| Use case | Complementary tool | Complexity | Estimated ROI |
|---|---|---|---|
| Automated QA testing | Playwright CLI | Low | High |
| Scraping with authenticated session | Persistent Chrome profile | Medium | Medium |
| Lead research + contact finding | Playwright | Low | High |
| Automated community engagement | Playwright + cron | Medium | Variable |
| Support ticket triage | Webhooks + Playwright | High | Very high |
For recurring tasks, you turn the script into a "skill": a short command you call whenever you need it. The process becomes consistent and repeatable. A Playwright CLI token consumes far fewer tokens than an MCP Chrome DevTools server, which matters as soon as you scale.
"Routines": cloud agents that run without you
Claude Code 2.0 introduced a feature that most people haven't fully grasped yet. Routines are packaged agents that run on Anthropic's infrastructure. Your computer can be shut down or you can be traveling: the agent runs regardless.
Three trigger modes are available. A cron schedule ("every day at 8 AM, generate an AI briefing in Notion"). A webhook ("when a form is submitted, research the lead, draft a personalized email, add to the CRM database"). A GitHub event (PR opened, issue labeled, commit pushed).
The webhook mode is what changes the game for agencies and product teams.
A webhook + a Claude routine = a fully automated lead qualification pipeline, no n8n, no Make, no additional stack. Jack Roberts demonstrated this live: a filled-out form triggers research on the prospect's website, a bespoke email draft, and the creation of a Notion record, all within minutes.
The main limitation is real: routines are probabilistic, not deterministic. The agent won't always do exactly the same thing. For strict workflows ("if X then Y, no exceptions"), stick with n8n or Make. But for anything requiring judgment (qualifying, summarizing, prioritizing, drafting), Claude Routines is hard to beat.
The agency model that Claude Code makes viable
This is where things get concrete from a business standpoint. Liam Ottley documented this with Tyler, a participant in his accelerator program: he visits clients on-site, sets up a Claude Code workspace connected to all their tools, builds the first automation live, then charges $2,500 USD/month as a retainer to deliver one or two new automations per month.
This model wasn't profitable in 2023. Development time was too long, APIs were too unstable, the back-and-forth was too costly. Today, an automation that would have taken a week of work ships in a few hours. The ratio between perceived client value and production cost flips completely.
For the client, the math is clear: $2,500 per month for someone who automates concrete tasks is often cheaper than a marketing agency subscription, with visible ROI from the very first meeting. Four clients at that rate, and you're past $10,000 per month with a one-person team.
I believe this opportunity is real, but only for those who know how to build robust systems. The combination of experienced developers + AI + strict processes is exactly what holds up over time, not prompts thrown together on the fly.
What separates a gadget from a system
On Reddit, the debate is raging. A full-stack developer with seven years of experience admits to "losing his skills" as he delegates more and more to Claude Code. A CTO questions whether to ban AI tools on his team after watching his senior dev ship barely maintainable code. The community is split: the problem isn't the tool, it's how you manage it.
An AI agent without clear architecture becomes unmanageable within weeks.
Files like CLAUDE.md, ARCHITECTURE.md, or CURRENT_STATE.md aren't luxuries. They are the project's memory. Without them, every session starts from scratch, code diverges, and technical debt accumulates silently. This is exactly the scenario I describe in the most common mistakes web developers make with Claude Code.
Breaking projects into short blocks with precise acceptance criteria, testing in a real browser rather than just in generation, documenting every architecture decision: that's what turns the tool into an industrial production system. Without it, you get the "good enough" that several devs on Reddit call out, code that works until the day it doesn't.
The real value doesn't come from the prompt. It comes from the system you build around it.
Verdict: a multiplier, not a shortcut
Automation with Claude Code isn't a shortcut. It's a multiplier. If your process is fuzzy, it amplifies the chaos. If your process is solid, it compresses delivery time dramatically.
The three most immediately actionable use cases: Playwright for anything involving a browser, routines for workflows triggered by webhooks or schedules, and the retainer model for monetizing these skills in B2B. These aren't promises: they are documented, tested configurations used by people who make a living from them.
What I see in the developers who actually extract real value from it: they don't use it to "code less." They use it to ship things they would never have had the time to build otherwise. That distinction makes all the difference, and it's exactly what I explore in the article on AI agents versus developers.
Vidéos YouTube
- Claude Code + Playwright Automates Literally Anything · Nate Herk | AI Automation
- How to Make $20,000/Month with Claude Code · Liam Ottley
- Claude Code 2.0 Is Here... Automate Anything · Jack Roberts
