AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Most people still use Claude as a blank chat box: open a new conversation, re-explain who they are, paste the same document again, copy the answer out by hand, close the tab. It feels fine because each answer is decent. But nothing carries over, so every session starts from zero and you pay the setup cost every time.
The rule: set up the five features in order, from who you are, to what you're working on, to where your files live, to how a task should be done, and give each one a single job.
| You need | Notes |
|---|---|
| A Claude account | Free works for most of this. Free accounts get up to five projects and one custom connector. |
| A paid plan (optional) | Pro is listed at $20 a month, or $17 a month billed annually, at time of writing. Claude's pricing page shows US dollars; check the amount in pounds, including VAT, at checkout before you rely on it. Pro adds unlimited projects and chat search. |
| About a day | Most of that goes on writing and trimming, not clicking. |
| Your own reference material | A short note about your role, two or three documents you reuse, and one task you repeat every week. |
| Slot | Feature | What it replaces | What skipping it costs you |
|---|---|---|---|
| Morning | 1. Instructions for Claude | Retyping "I'm a… keep it short…" in every chat | Generic answers pitched at nobody |
| Morning | 2. Memory | Re-explaining ongoing context | Claude forgets what you told it last week |
| Late morning | 3. Projects | Pasting the same files into new chats | Wrong or stale versions of your documents |
| Afternoon | 4. Connectors | Copying text in from Drive, Gmail and the rest | Claude only sees what you remember to paste |
| Afternoon | 5. Skills | Long "do it like this" prompts | The same task done slightly differently each time |
Do steps 1 and 2 together. Instructions are what you tell Claude on purpose; memory is what Claude picks up from your chats.
If you would rather draft it with Claude, use this prompt. Fill in the square brackets.
You are helping me write the account-wide instructions that Claude will read at the start of every conversation.
About me:
- Role and sector: [ROLE_AND_SECTOR]
- Who I usually write for: [AUDIENCE]
- Formats I want by default: [FORMATS, e.g. short bullet points, British English, £]
- Things I never want: [DISLIKES, e.g. hype words, invented statistics]
Steps:
1. If any field above is blank or vague, ask me about it before drafting. Do not guess.
2. Draft no more than 10 short lines in the second person ("You are helping…").
3. Leave out anything that only applies to one client, project or task.
4. After the draft, list anything you left out and tell me whether it belongs in a project or a skill instead.
Before you answer, check that each line would hold true for every chat I have, and remove any line that would not.
Two controls are worth knowing. An incognito chat (the ghost icon when you start a new chat) neither uses nor creates memories. You can also switch memory off for a single chat from the + menu, but only before you send your first message.
Start with one project, not ten. Each chat inside a project is separate: something decided in one chat is not available in the next unless you add it to the project knowledge.
Connect the one or two services you copy from most often (usually your documents and your email), then stop there. Every connector you add gives Claude more data to read and, for some services, the ability to take actions. Custom connectors that use a remote MCP server are added under Customize > Connectors > Add > Custom > Web. Free accounts are limited to one.
Use this prompt to turn your weekly task into the text of a skill:
You are helping me turn a task I repeat into a reusable Claude skill.
The task: [TASK_NAME]
How often I do it: [FREQUENCY]
A good past example (pasted below or attached): [EXAMPLE_OUTPUT]
Steps I always follow: [MY_STEPS]
What usually goes wrong: [COMMON_FAILURES]
Steps:
1. If the example or the steps are missing, ask for them. Do not invent a process.
2. Write a short description saying when this skill should be used and when it should not.
3. Write the instructions as numbered steps, including the exact output format.
4. Add a final checklist the skill runs through before it hands over its answer.
5. Tell me which files the skill folder needs, and remind me to check Anthropic's current skill format before I zip and upload it.
Before you answer, check that someone who had never seen my example could follow your steps and get the same shape of result.
The trap is putting everything into one place, usually the account-wide instructions. Client rules, file summaries and task steps all end up in a box that applies to every chat, and the result is slower, muddier answers. Keep each layer to its job: instructions for who you are, memory for what changes, projects for one body of work, connectors for live files, skills for repeated methods.