AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
Most people spend their first week with Claude treating it like a search bar: a fresh chat, a one-line question, a generic answer, repeat. It feels productive because every reply is fluent, but Claude starts from zero each time, knows nothing about your work and never touches the files and apps where that work lives. The fix is not cleverer questions. It is four pieces of setup that take about an evening.
The rule: give Claude somewhere to work (a project), something to work with (a connector), a way to repeat good work (a skill), and permission to ask before it guesses.
| What you need | Notes at time of writing |
|---|---|
| A Claude account at claude.ai | Free works for most of this guide |
| Free plan | Projects (up to five), web connectors, skills and memory are all available. Free is limited to one custom connector; directory connectors are not described as limited |
| Pro plan | Listed at US$20 a month on Anthropic's pricing page, roughly £15 before VAT at today's exchange rate. Anthropic does not publish a fixed £ price there, so check the figure at checkout |
| One real piece of work | A document, a recurring task, or an app you use daily (Gmail, Google Drive, Notion and so on) |
Paid plans add past-chat search and a larger project capacity; you do not need either tonight.
In the message box, the + button is where you attach files, reach Connectors, and switch per-chat options such as memory for that conversation. The model selector next to it chooses which Claude model answers. The sidebar holds your chats, Projects and Customize, which is where connectors and skills now live.
Start by attaching something real and asking for a prioritised reading of it, not a summary:
You are helping me make decisions about [WHAT I ATTACHED, e.g. "our Q3 supplier contract"].
Context: I am [MY ROLE] and I need to decide [THE DECISION THIS FEEDS].
1. Read the whole attachment before answering. If any part failed to load or is unreadable, say so first.
2. Pick the three points I most need to act on, ranked by consequence if I ignore them.
3. For each: one sentence on what it says, one on why it matters to my decision, and the section or page it comes from.
4. List anything ambiguous that I should check with a person rather than rely on you for.
Format: a numbered list, then a short "Check with someone" list. No preamble.
If you cannot tell what decision I am making, ask me before you rank anything.
Before replying, confirm each point cites a real location in the document.
Fill in what you attached, your role and the decision it feeds.
A connector sees only what your account can see in that app, and it can act there, not just read. Then ask what it is good for in your work:
You now have access to [APP NAME] through a connector.
I run [MY KIND OF WORK, e.g. "a two-person bookkeeping practice"], and my most repetitive tasks are [TWO OR THREE TASKS].
1. Tell me which actions you can actually perform in [APP NAME] through this connector, and which you cannot. Do not guess: if you are unsure whether an action is supported, say so.
2. Suggest five specific uses tied to my tasks above, each as a one-line example request I could paste in.
3. Mark any suggestion that would change, send or delete something, so I know to review it first.
Format: a short "can / cannot" list, then the five suggestions as a numbered list.
If my tasks are too vague to suggest anything specific, ask me one question first.
Fill in the app name, your kind of work and your repetitive tasks.
One trap: chats inside a project do not share context with each other unless the information is in project knowledge. If a decision matters, add it as a file. Each project also keeps its own separate memory.
PROJECT INSTRUCTIONS
Purpose: this project covers [WHAT THE PROJECT IS FOR].
About me: [ONE LINE ON WHO I AM AND WHAT I DO].
Audience for outputs: [WHO READS WHAT YOU PRODUCE].
Working rules:
- Use British English. Keep replies brief unless I ask for depth.
- Treat files in project knowledge as the source of truth; if they conflict with what I say in chat, point out the conflict.
- Before any substantial task (a draft, a plan, an analysis), give me your approach in no more than three bullets and wait for my approval.
- If an input you need is missing, ask for it instead of inventing it.
- Never present a figure, date or name as fact unless it comes from my files or I gave it to you.
Fill in the purpose, one line about you and who reads the outputs.
A skill is a packaged set of instructions Claude loads when a task matches it. Skills need code execution switched on: on Free, Pro and Max that is Settings > Capabilities > Code execution and file creation. On Team and Enterprise an owner enables skills in organisation settings first.
When a chat has gone well, ask Claude to package it:
The process we just followed worked. Package it as a reusable skill.
Name: [SKILL NAME]
Use it when: [THE SITUATION THAT SHOULD TRIGGER IT].
1. Write the steps in the order we actually followed them, so you would repeat them the same way next time.
2. Include the preferences I stated in this chat (tone, format, what to leave out), quoted or paraphrased accurately.
3. Write a one-sentence description that makes clear when this skill should and should not be used.
4. Do not add steps we did not take.
Show me the full skill for review before you save or package anything, and ask if any step is unclear.
Fill in the skill name and when it should be used. Save the file Claude produces, then manage it under Customize > Skills, where you can toggle each skill on or off or upload one as a ZIP via + > Create skill > Upload a skill. Some older Anthropic tutorials still show the path as Settings > Capabilities > Skills; if your screen differs, look in both.
For anything where a wrong assumption is costly, add this to the end of your request:
Before answering, ask me the questions you need answered to give advice that fits my situation, up to [NUMBER, e.g. five] at a time. Keep asking in rounds until you have what you need, then tell me what you are still assuming before you give your answer.
Fill in how many questions per round you can tolerate. Skip the old "until you are 95% confident" wording: the model has no calibrated confidence meter, so the number is theatre. Asking it to list remaining assumptions is something you can actually check.