Video

Claude AI for Beginners: A Six-Level Guide from First Prompt to Claude Code

Published 11 min read
TL;DR

A walkthrough of six levels of using Anthropic's Claude, from three-part prompts and file uploads to projects, memory, connectors, scheduled Cowork tasks and Claude Code, using a candle-shop example.

Why it matters

It gives you a clear order for learning Claude, plus the plan limits and safety settings to check before you hand it files, apps or autonomous tasks.

Metics Media · Watch on YouTube · Loads from youtube-nocookie.com only after you press play

Claude is Anthropic's AI assistant, and the quickest way to get real value from it is to climb in stages: write better prompts, give it your files, make it build things, give it memory, connect your apps, then let it work on its own. Matt of Metics Media organises a beginner tutorial into six such levels and runs every step on one demo business, a small handmade candle shop called Amberlight. He asks viewers not to copy that example but to run each prompt on their own topic, such as a newsletter, a job search or schoolwork.

The interface and model names in the demo reflect the software as recorded and may have changed since. Matt says so himself, and advises following the concept rather than the exact screenshots.

The six levels at a glance

Level What you do Plan needed (per the video)
1. Talk to Claude Write structured prompts, refine answers, pick a model Free
2. Give it your world Upload files and images, web search, Research Research needs a paid plan
3. Make real things Excel, Word and PowerPoint files; interactive artefacts Shown in the video after the upgrade
4. Make it remember you Projects, memory, personal instructions, mobile and desktop apps Not specified in the video
5. Connect your apps Connectors, custom connectors, skills Free plan allows one custom connector
6. Work while you are away Cowork, scheduled tasks, Claude Code Cowork needs a paid plan

Level 1: Prompting and choosing a model

A prompt is simply the message you send to Claude. Matt's core advice is the three-part prompt: context, task and format.

  • Context: who you are and what you are working with.
  • Task: exactly what you want.
  • Format: length, tone and structure.

His demonstration compares two prompts. The weak one, 'Write a product description for a candle,' returned a generic result. The stronger one named the shop, the soy wax and the natural scents, the customers' values, the product (an Autumn Cabin candle), a warm tone, a limit of under 100 words and one short line that makes you want to light the candle. That produced a noticeably more personal description.

Two habits follow from this:

  1. Expect different answers each time. Matt stresses that the same prompt will not return his exact output. Copy the pattern, not the result.
  2. Reply and refine instead of starting over. Claude keeps the whole conversation in view, so a follow-up such as asking for half the length while keeping the 'hand poured' mention reshapes the draft. Matt treats every answer as a first draft.

Models, effort and limits

The model names change often, so Matt gives a pattern instead: a fast model for everyday tasks and a more powerful model for complex work. At the time of recording he listed Sonnet as the default, Haiku as lighter, Opus as a step up and a model called Fable as the top one (some only on paid plans). Move up for long contracts, important writing or complex planning, and down when you are firing off many small questions. You can switch mid-conversation.

An effort setting controls how long Claude thinks before answering. Raise it for multi-step problems where details must be right. Higher effort and stronger models use your usage allowance faster.

Every plan, free or paid, has usage that refills on a schedule. Hitting the limit breaks nothing; you wait for the reset or upgrade. Matt's advice is to start free and upgrade on the day the limit starts to annoy you.

Account and pricing basics

Claude is free to start, and Matt says most of what he shows works on the free plan. The cheapest paid plan, Pro, is $17 a month billed annually or $20 a month billed monthly, as quoted in the video. During onboarding there is a choice about whether your chats can be used to train Anthropic's models. Matt calls it a personal preference and turns it off for privacy.

Level 2: Files, images, web search and Research

Start a new chat for each new topic. Matt's rule is same task, same chat; new topic, new chat. Separate chats keep Claude focused and use less of your allowance than one endless conversation. You can pin frequently used chats.

Files and images. Claude reads PDFs, Word documents, spreadsheets and images. In the demo, a supplier price list PDF was summarised in a table and Claude identified the cheapest option per kilogram (a bulk format of a product Matt calls Golden Soy 464). A product photo drew critique that the background was too present and read as a casual snapshot rather than a product shot. Matt's challenge: upload the most boring document you have and ask Claude to explain it.

Web search. Claude does not automatically know live news or prices. Check that web search is ticked in the plus menu, then ask. The answer shows its sources, and Matt's advice for any AI is to look at the sources when facts matter.

Research. This feature reads dozens or even hundreds of sources and returns a cited report. It needs a paid plan, and it uses your allowance faster than normal chat, so Matt saves it for questions that deserve it. His prompt asked about the handmade-candle market: who buys, what they pay, how small brands market, and the three biggest mistakes new candle businesses make. The result was a document with a short summary at the top, key findings and a source list. Matt calls it one of the two features that made upgrading worthwhile for him.

Level 3: Files and artefacts

Claude can produce two kinds of output outside the chat:

  • Downloadable files: Excel, Word, PowerPoint and PDF. The demo inventory spreadsheet had a calculated profit-margin column and eight sample candles; Matt says it contained working formulas and opened in Excel.
  • Artefacts: interactive pages and mini apps built in a panel beside the chat. A landing-page request produced an HTML page with a warm colour scheme and a candle flicker animation, using placeholder images. Follow-up requests made the background cream, enlarged the headline and then turned the page into a scent quiz that recommended a candle.

Matt's challenge for this level is to ask Claude to build one small tool for your life, such as a budget calculator or workout planner.

Level 4: Projects, memory and the apps

Projects are workspaces where related chats share knowledge. Two features matter:

  • Knowledge: files added once are available in every chat in the project.
  • Instructions: describe your situation once and each chat inherits it.

After Matt added a supplier price list, the research report and short brand instructions, the prompt 'write three ideas for our first newsletter' produced Amberlight-specific ideas without any extra explanation. He calls this the biggest step from casual use to working with Claude.

Personal instructions and memory work at the account level. Personal instructions apply to all chats (Settings, General). Memory lets Claude notice and retain details from conversations; you can read it, edit it by chatting, or delete it. An incognito chat (ghost icon) is not saved, adds nothing to memory and is not used for training; in the demo, Claude could not say who the user was.

Importing from ChatGPT. In Settings, Memory, there is an option to import memory from other AI providers. Claude supplies a prompt to paste into the other assistant. In Matt's case ChatGPT did not return a file but pointed him to its memory summary in Settings, Personalization, Memory, Manage. He copied that text back into Claude and chose add to memory. Matt describes the import as experimental, so details may come through imperfectly, and they can be edited afterwards. He also recommends enabling the toggle that lets Claude search and reference past chats.

Mobile and desktop apps. The iPhone and Android app syncs your account, chats, projects and memory. Matt's suggested pattern is capture and continue: photograph something or talk through an idea on the phone, then pick the conversation up on your computer. In his demo, a photo of a candle returned a price estimate of about $48 to $65 for a premium handmade candle, though Claude miscounted the wicks (four instead of three). A desktop app for Mac and Windows is needed for the later levels.

Level 5: Connectors and skills

Connectors let Claude plug into apps such as Gmail, Google Drive, Slack and Google Calendar, read their contents and act in them. You choose what to connect and can disconnect at any time.

  1. In a chat, open the plus menu, then Connectors, Add connector, Browse connectors.
  2. Pick an app and authenticate with that service.
  3. Set the permissions Claude receives. Matt granted broad calendar access; you can grant less.
  4. Ask in plain language. His example, asking which afternoon next week had the most free time for deep work, returned a Wednesday afternoon recommendation without any copying or screenshots.

Custom connector example (sponsored)

This segment is sponsored by Omnisend, an email and SMS platform for e-commerce. Treat the demonstration as a sponsored showcase, not an independent comparison. Matt and the sponsor read-out make claims such as lower prices for switching stores, free migration of lists and automations within about five days, more than 150,000 stores and a 4.8 average Shopify rating; those are the sponsor's figures as relayed in the video and were not independently checked.

Matt added Omnisend as a custom connector using its MCP server address, which he provided in the video description. The free Claude plan allows only one custom connector. The connection did not complete the first time and needed a manual reconnect, which he describes as an occasional bug. Claude asks permission before each action, and you can permit actions automatically.

The prompt asked for a welcome email draft with subject line, preview text, a 10% first-order discount and brand story. Claude hit a text-corruption issue, diagnosed it and rebuilt the draft with safer characters, then created a draft campaign in Omnisend. Matt checked it in the dashboard: subject, sender, pre-header, merge tags for first names, buttons and a coupon were present. Nothing was sent; the draft awaited his review.

Skills

A skill is a text file of detailed instructions, a recipe card written for Claude, that you install once. In Customise, Browse, Matt installed Theme Factory, which supplies 10 themes for styling slides and documents. Typing a slash in a chat lists installed skills, and Claude may also use one automatically when the job fits. He asked for a short deck introducing Amberlight to a local store owner, chose the Golden Hour theme adapted to the earlier website, and got a polished presentation. Some icons looked odd, which he says can be redone on request.

Level 6: Cowork, scheduling and Claude Code

Cowork

Cowork (desktop app, paid plan) lets Claude carry out a task autonomously. Claude can only read and write in the folders you connect; the rest of the computer stays off limits. The approval modes:

Mode Behaviour
Manual Asks before each action
Automatically approve Works on its own with safety checks
Skip all approvals Never stops to ask

Matt recommends starting with manual approval and loosening it once you trust the workflow. He himself used automatic approval with Opus and high thinking effort for a folder of receipts: Claude extracted date, supplier and amount into one expense spreadsheet. A follow-up asked it to rename each receipt with date, supplier and reference number. Claude offered three options with a recommendation, then renamed the files and updated the spreadsheet to match. He notes that runs differ between users because Claude may take a different path to the same result.

Scheduled tasks

In Cowork, the Scheduled section lets you describe a task once and set how often it runs. Matt created a Monday 6:30 AM web search for candle and home-fragrance trends that produces a detailed report. His tips:

  • Run it once as a test so routine permission prompts get pre-approved.
  • Edit the task afterwards to change permissions, model, frequency or details.
  • Matt says scheduled tasks can run in the cloud, so they can happen when your computer is off.

Claude Code

Matt says Cowork and Claude Code run on the same engine: Cowork points it at everyday files and documents, Claude Code at building software. You still use plain language. He combined two tools:

  1. Asked Claude for a markdown summary of everything known about Amberlight as a design brief.
  2. Pasted it into Claude Design, chose the UI mock-up template and a top model at high effort, and answered clarifying questions. The result was a homepage, shop and about pages with a working scent quiz and cart, but design only, with placeholder images swapped for real photos on request.
  3. Used Share, then Send to Claude Code, and gave it a folder and the instruction to turn the design into a working website.

Claude Code's permission options are manual, accept edits, plan and auto. Matt uses auto but recommends a more restrictive mode on your first use.

What worked and what did not. The build took about 20 minutes. The images initially did not come through and the email service was not wired up. After a follow-up, Claude Code added the photos (converting them to smaller JPEGs on its own judgement) and connected the signup form to the Omnisend connector. Matt still needed an API key from Omnisend to finish that link. He says the site is a real working site running locally but lacks personalisation and an e-commerce service such as Shopify.

A reusable tool. In a second Claude Code demo, Claude built a desktop app that reads an inventory spreadsheet and produces a print-ready PDF of product labels. After Matt edited the spreadsheet (renaming one product and adding another), re-running the app produced updated labels without any further conversation. His point: Claude can build a tool that keeps doing the job without Claude, and without using up usage limits.

Takeaways

  • The skill underneath every level is the same: give context, give a task, refine the result.
  • Match model and effort to the stakes, and watch usage, since Research, stronger models and high effort use allowance faster.
  • Check sources whenever facts matter.
  • Give permissions gradually: limit connector scopes, connect only specific folders, and start autonomous work in manual approval.
  • Every result in the video is one run on one demo business. Matt says repeatedly that your outputs will differ.

Primary sources

Checked claims

ClaimStatusEvidence
A prompt is simply the message sent to Claude or another AI.confirmedIn the video at 4:01: “A prompt is just the message you send to Claude or any other AI.”
The recommended prompt structure is context, task and format.confirmedIn the video at 4:33: “So watch what happens with what I call the three-part prompt. That includes context, task, and format.”
Use one chat per task and a new chat for each new topic.confirmedIn the video at 9:56: “The rule is simple. Same task, same chat. New topic, new chat.”
Claude is free to start and most demonstrated features work on the free plan.confirmedIn the video at 2:25: “Claude is free to start, and most of what I'll show you works on the free plan.”
New users should begin Cowork with manual approval.confirmedIn the video at 41:14: “Start with manual approval.”
Cowork and Claude Code share the same underlying engine, aimed at different work.confirmedIn the video at 46:59: “Cowork and Claude Code actually run on the same engine.”

Found a mistake? Use Report above; an editor reads every one. Editorial policy

Discussion · 0

No comments yet. Start the discussion.

Sign in to join the discussion.