AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
Most people "learn Claude" by installing everything at once: the app, a dozen connectors, a folder of shared skills, and Claude Code because a thread said so. It feels productive. Then something misbehaves, you can't tell which layer caused it, and you drift back to pasting text into a browser tab.
The rule: add one layer at a time, and don't add the next until the current one has done a real job for you.
The order: desktop app, connectors, skills, a real folder in Cowork or Claude Code, then a standing instructions file. A weekend covers all five.
| You need | Why | Cost at time of writing |
|---|---|---|
| A Claude account | Everything below signs in with it | Free plan exists |
| A paid plan (Pro or above) for Steps 4 and 5 | Cowork and Claude Code are not on the free plan | Pro is listed at US$20 a month, or US$17 a month billed annually; the UK price shows in pounds at checkout and changes with the exchange rate and VAT, so check claude.ai/upgrade before you buy |
| macOS 11+, Windows 10+, or Ubuntu 22.04+/Debian 12+ | Supported desktop systems | Free |
| One real job per step | So you can tell whether each layer earned its place | Your time |
You're going to be my working assistant for the next few weeks. Before you help with anything, interview me.
My role: [YOUR_JOB_TITLE_AND_SECTOR]
The three tasks that eat most of my week: [TASK_1], [TASK_2], [TASK_3]
Tools I use daily: [EMAIL_APP, CALENDAR, FILE_STORAGE, OTHERS]
Steps:
1. Ask me up to five short questions, one at a time, about how I work and what "done well" means for each task.
2. Write a one-page working brief: priorities, preferred tone, and for each task where you can save time and where you should stay out.
3. Mark anything you inferred rather than heard from me as "(assumed)".
Don't invent details about my organisation. If an answer is missing, ask. Before you give me the brief, check it against my answers and remove anything I didn't say.
Fill in your role, three heaviest tasks and daily tools. Keep the brief; Step 5 reuses it.
Connectors let Claude read from (and sometimes act in) services such as Gmail, Google Drive, Google Calendar, Slack, Microsoft 365 and GitHub. Anthropic says web connectors are available on every plan, including free.
/).On Team and Enterprise plans, an owner must enable a connector for the organisation first. Start with two connectors, not ten, and test them on a real job:
Using my connected [EMAIL_SERVICE] and [CALENDAR_SERVICE], help me prepare for [DAY_OR_DATE].
1. List my meetings that day with time, attendees and the stated purpose.
2. For each meeting, find the most recent email thread with those attendees from the last [NUMBER] days and summarise it in two lines.
3. Flag any meeting where I seem to owe someone a reply or a document.
Output a table: Meeting | Time | Context | What I owe.
Only use what you actually find in my accounts. If a meeting has no matching thread, write "nothing found" rather than guessing. Don't send, draft or change anything. Before answering, confirm every row comes from a real email or calendar entry.
Fill in your email and calendar service, the day, and how far back to search.
Skills are packaged instructions (and sometimes scripts) that Claude loads when a task calls for them. Anthropic ships built-in skills for Excel, Word, PowerPoint and PDF files. Skills need code execution switched on.
Only add a skill for a job you've done by hand at least twice. Anthropic advises installing skills only from trusted sources and reading the files first, because a malicious skill can attempt prompt injection or leak data. To draft one from work you repeat:
I want to turn a task I repeat into a reusable Claude skill.
The task: [DESCRIBE_THE_REPEATED_TASK]
An example of good output: [PASTE_ONE_REAL_EXAMPLE]
Things that always go wrong: [COMMON_MISTAKES]
1. Ask me what's missing before you start: inputs, audience, format, length.
2. Write a SKILL.md with a short name, a one-sentence description of when to use it, and numbered instructions.
3. Include a checklist Claude runs before finishing the task.
4. Keep it under one page. No scripts unless I ask.
Show me the SKILL.md and a list of anything you assumed.
Fill in the task, paste one good example, and list the mistakes you keep correcting.
Here Claude stops being a chat box and works on your files. Two routes in the desktop app, both paid-only:
| Route | Where | Best for |
|---|---|---|
| Cowork | Select Cowork in the message box, then describe the task | Documents, spreadsheets, research, file tidying |
| Claude Code | Click the Code tab, choose Local, pick a project folder, model and permission mode | Anything involving scripts, code, or many files changed at once |
Some accounts have a unified message box with no separate Cowork option, where any conversation can take Cowork-style tasks. If you prefer the terminal, the official installer is curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux and WSL, or irm https://claude.ai/install.ps1 | iex in Windows PowerShell; then run claude inside a folder and log in.
Claude Code isn't only for programmers: pointed at a folder of invoices, notes or exports, it can rename, sort, summarise and build reports. Work on a copy first:
You're working in a copy of my [FOLDER_DESCRIPTION] folder. Goal: [WHAT_I_WANT_AT_THE_END].
1. Before changing anything, list what's in the folder and propose a plan as numbered steps.
2. Wait for me to approve the plan.
3. Then carry it out, and keep a log file called changes-log.md listing every file you created, renamed or moved.
4. Don't delete anything. If something looks like a duplicate, list it in the log instead.
If the folder contains something you don't understand, ask me rather than guess. When done, check the log matches the folder and tell me anything you skipped.
Fill in what the folder holds and the result you want, such as "a summary spreadsheet of every invoice by supplier and month".
Each Claude Code session starts with a fresh context. A CLAUDE.md file carries your instructions across sessions: ~/.claude/CLAUDE.md for you across all projects, or ./CLAUDE.md inside a project folder. Running /init in a session generates a starting file from what Claude finds in the folder; if one already exists, it suggests improvements instead of overwriting it. Run /context to confirm the file loaded. Add your Step 1 brief, trimmed to rules Claude couldn't work out itself.
changes-log.md, and every change in it is one you approved./context lists your CLAUDE.md under Memory files.Shift+Tab in the terminal) while you're learning.