Anthropic’s coding agent runs in your terminal as text, and can possibly build your next startup in an afternoon. Let’s explore...
Pics/iStock, Claude
Claude Code is Anthropic’s coding agent that runs in a terminal and is designed to help you build, debug, and automate work inside real coding projects. In the last month or so, since it’s been available to the public, everyone from prominent developers from competing AI companies to average Joes with no coding experience have been praising Claude Code. The whole thing runs from PowerShell on Windows or Terminal on Mac and Linux machines. It is very simple to use and very useful; in fact, Anthropic itself has been using it to deploy some new products.
How does it work?
Claude works like any LLM except that it is designed to deliver a complete application or program. So you essentially describe what you want to build and all the features it needs to have and Claude code will build it. The more detailed your prompt is, the better your result. I personally like to take my idea to ChatGPT or Claude and explain what I want to build, then ask it to produce an action plan. I use this action plan to build my product.
How to get started
You need a Claude account to get started. A working computer and Internet. Claude works on Windows, Mac, and Linux.
On Mac:
. Open a terminal window and type this:
. curl -fsSL https://claude.ai/install.sh | bash
For Windows:
. Open PowerShell and type this: irm https://claude.ai/install.ps1 | iex
. That’s it, it’s installed.
In the Terminal or PowerShell, head on to the directory that contains your code or where you want to start a new project, and type.
Claude
That’s it, just follow the instructions and connect it to your account, and you can now start building an app.
First steps
Once it’s running, you can use it in a few practical ways. If you want to add something new, for example, a small feature on a website or a new button in an app, ask Claude Code to first outline a simple plan, then follow that plan step by step. This keeps things organised and easier to understand.
If something is broken, you can paste the error message or describe what’s going wrong. Claude Code can often point out the likely cause and suggest what to change to fix it.
It’s also useful as a map. You can ask simple questions like “Where does this part live?” or “What happens when I click this?” and it can guide you through the project’s files in plain language, helping you understand it.
Finally, it can help with repetitive tasks, like running checks, cleaning up files, or preparing updates automatically.
For your first project, choose something small but complete: add one new feature, update the instructions, and make sure it still works, or look for a practice project online.
Pic/iStock
Tips and tricks
Ralph Wiggum, named after a character on The Simpsons, it is a nickname for a simple AI-coding workflow: you wrap Claude Code in a Bash loop that keeps re-running the same task until it’s done or you hit a safety limit. Each pass sees the repo’s updated files and git history, so the model can review its own changes, notice what’s broken, and fix it – like Ralph’s simplistic persistence in the show, but productive. It shines for mechanical, well-defined jobs with automated checks (tests, type checks, linters) and clear “done” criteria. However, it can fall apart with vague goals, subjective quality, security-sensitive code, not to mention runaway token costs, so max-iterations and human review are essential. In Claude Code, a plugin uses a “Stop hook” to block the assistant from exiting, then re-feeds the original prompt for another round. The point isn’t a first prompt; it’s repeated improvement. The idea can be applied to other tools by looping until checks pass. Ralph Wiggum is already available on Claude, and you can initialise it directly. Get shit done.
Getshitdone (GSD) is a lightweight “spec-driven development” toolkit designed to make Claude Code more reliable by adding a structured workflow on top of it. Instead of ad-hoc vibecoding, it guides you through capturing a clear idea (goals, constraints, edge cases), then turns that into written specs and a step-by-step roadmap Claude can follow. You install it with a single command (npx get-shit-done-cc) and then use simple slash commands (like /gsd:new-project) to generate key files such as PROJECT.md, ROADMAP.md, and a persistent STATE.md “memory” file that keeps sanity for the project. The system breaks work into phases and small atomic tasks, often running implementation in fresh sub-agent contexts to avoid long-session drift. It also recommends running Claude Code in a low-friction permissions mode (or configuring granular allowlists) so automation isn’t constantly interrupted. In short: fewer vibes, more shipped features. You can see how to run it from its GitHub (https://github.com/sabalioglu/getshitdone)
Skip permissions
A little risky, but you can skip the need to continuously intervene with this command: claude--dangerously-skip-permissions when starting Claude.
Conclusion
Claude Code isn’t magic, and it won’t replace good judgment, but it will remove much of the friction that keeps ideas from becoming real things. Used well, it’s like having a patient helper who can draft, organise, and tidy up code while you focus on what you actually want to build. Start small, keep your instructions clear, and let it earn your trust one task at a time. And if you’re tempted to run it unattended overnight, remember: automation is brilliant right up until it isn’t and you end up burning a bunch of credits. Set limits, use checks, and review what it changes.
The goal isn’t to hand over the steering wheel; it’s to arrive faster, with fewer bumps in the road.
Subscribe today by clicking the link and stay updated with the latest news!" Click here!
