Claude Code for Beginners: Your First Week

Claude Code for beginners is Anthropic's terminal coding agent: you run it in a project folder, you describe a job, and it edits files. You do not need to be a senior engineer to start. You need a folder and a willingness to read the diff.

Most beginner posts skip the part that hurts. Claude Code cannot see your screen unless you give it context, and a pasted PNG is a weak way to do that.

I would rather you have a boring first week than a clever one. One repo. One feature. Named feedback.

The install and command basics live in how to use Claude Code. This page is the first-week method.

What is Claude Code, if I am new?

Claude Code is an agent that lives in the terminal. It reads your project, proposes edits, and runs commands you allow.

It is not an IDE. If you want a sidebar chat on a file tree, that is Cursor. Both can be beginners' tools. They feel different.

You bring the model through your Anthropic setup. PinVari and other local tools should not replace that. They only add context.

A longer Claude Code tutorial exists if you want command-by-command. Stay here if you want the week plan.

How do I install Claude Code and not get stuck?

Install the CLI the official way, then open a real folder. Do not start in your home directory.

If you already have PinVari and want screen context, connect MCP with a scoped command. The bare form errors.

claude mcp add --scope user pinvari -- "$HOME/.pinvari/mcp/pinvari-mcp"

The app must be installed and running. The connector talks to 127.0.0.1:3402. One-click connect inside PinVari → Connect → Claude Code does the same job.

Claude Code MCP is the deeper explainer. You can skip it until day four.

Claude Code for beginners: the seven-day loop

Day one: run Claude Code in an empty project. Ask it to scaffold the list app. Accept the files. Run whatever start command it prints.

Day two: click the UI. Write three outcome sentences. "Empty state should say there are no names." Do not mention frameworks.

Day three: try a screenshot once so you feel the pain. Then read Claude Code screenshots and stop making that your default.

Day four: connect local MCP. Hold ⌥⌘A, circle the broken control, speak the outcome. Ask Claude for pinvari_next_instruction.

Day five: let it request a capture. pinvari_request_capture lights the notch island. You point. It continues.

Day six: break something on purpose. A long name. A second display if you have one. File what you see with a named element.

Day seven: start a second small job in the same repo. Resist a new tool.

Terminal comfort

Goal: You can cd into the project and start the agent.

Not the goal: Memorizing flags.

Check: You can run the app without asking the agent where you are.

Diff comfort

Goal: You notice extra files.

Not the goal: You could have written the patch.

Check: You reject one change this week.

Context comfort

Goal: You give a named control, not a paragraph.

Not the goal: You become an MCP author.

Check: One capture becomes one correct edit.

Why beginners think Claude Code "does not see the UI"

Because it does not. It sees the repo and whatever you put in the prompt.

A screenshot helps a little and costs a lot. The model still guesses the element. Wrong-element edits are a genre at this point.

macOS can name the element under a point. Role, label, frame, confidence, circled vs dwelled provenance. Full window text up to 40,000 characters, including scrolled-out text. Browser URL from AXWebArea.

On-device transcription and OCR. Nothing uploaded by default. You still send the bundle to Claude Code because you asked it to work.

Vague captures should come back UNSPECIFIC. Ask, never guess. If two Saves exist, say which group they live in.

Deixis: "this" binds to the pointer at the moment you said it. Dwell ~0.2s counts. Three marks in one breath get three word buckets.

Beginner mistakes I would budget for

Starting in a huge cloned repo. The agent will thrash. Use a small folder.

Accepting every edit. You will not learn what a blast radius looks like.

Using CleanShot or ⇧⌘5 and calling it context. Pixels are not a name.

Chatting in Slack about the bug, then paraphrasing for Claude. Circle while the build is open.

Ignoring multi-display. If the preview is on the studio display, the mark must remember that monitor.

Paying for another voice subscription when you already need words bound to a pointer. Wispr Flow is fine for dictation. It is not an element resolver.

What "done" looks like after a week

You can start Claude Code in a folder without a guide.

You can describe a user job in two sentences.

You can reject a diff.

You can hand it a named element at least once.

You know pinvari_mark_done exists, even if you only used pinvari_next_instruction.

If you want the capture side on a Mac, PinVari's launch price is $39 (first 500 licenses, then $59) via Polar. macOS 14+, Apple Silicon and Intel, notarized Developer-ID DMG. Team seats available. No subscription for the core app.

Public tool docs: pinvari.com/mcp. Connect docs: pinvari.com/connect.

What should a Claude Code for beginners prompt look like?

Short. Outcome first. Files if you know them. Element if the job is visual.

Bad: "Improve the UI and make it production ready." The agent will invent a redesign.

Better: "Add stays disabled when the name field is empty. Do not change styles. Test by typing one letter and watching Add."

Better still: send a capture of Add with the spoken line, then "only this control."

Put the check in the prompt. If you cannot say how you will know it worked, you are not ready to run the agent.

A second prompt shape I like: "List the files you will touch, then wait." Beginners skip that and then cannot name the blast radius.

A third: "If you are not sure which Save, ask. Call request-capture if you have it." That one sentence prevents a class of wrong-element edits.

Do not paste your entire README every turn. Point at the folder by being in it. Claude Code already has the repo.

Do not paste five blog posts as "context." One named control beats a thousand tokens of prose.

If you need a template, keep three bullets on a note: outcome, out-of-scope, check. Reuse it until it is muscle memory.

Claude Code for beginners is mostly that template plus the courage to reject a diff. The CLI is the easy part.

FAQ

Is Claude Code good for beginners?

Yes, if you accept the terminal and you keep the first project small. Claude Code for beginners is a bad fit for "build everything" prompts.

If the terminal is a hard stop, use Cursor for a week, then come back. The agent ideas transfer.

Do I need to know the terminal first?

You need cd, ls, and the courage to paste a start command. That is enough for day one.

You will pick up more commands because the agent runs them. Read what it runs.

How is Claude Code different from Cursor for a beginner?

Claude Code is a terminal agent. Cursor is an editor with an agent.

Cursor is easier to look at. Claude Code is closer to repo truth and scripts. Both can connect to the same local MCP.

Can Claude Code see my screen?

Not by itself. You paste pixels, or you connect a local screen MCP.

A resolved accessibility element is the strong form. A full-window PNG is the weak form.

What should I do when Claude Code edits the wrong button?

Stop and give it a name. Circle the control, speak the outcome, send the capture.

If confidence is low, do not let it guess. Tell it there are two similar controls and which parent group you mean.

Is there a free way to try this workflow?

Claude Code's own billing is separate. The Mac capture app I use is a one-time license, not a monthly seat.

You can also start with built-in screenshots for a day so you feel why they fail. Then add names.