AI Guides › Step-by-step guides

The Claude Four-Part Setup: Memory, Skills, Connectors and an Ask-First Rule

By Nigel Guy · 7 min read

Most people open Claude, type a request, get a plausible answer and then spend the next ten minutes correcting it for facts they have already told it, files it cannot see and assumptions it made silently. The default feels fine because every single reply is fluent. The cost is spread across every conversation.

The fix is a one-off setup with four parts, each aimed at one failure: Claude forgets you, doesn't know your routines, can't see your real material, and guesses when it should ask. This guide takes about twenty minutes. It does not turn Claude into something it isn't, and the guardrails at the end say where it still needs you.

The rule: give Claude durable context once (memory, skills, connected data), then give it a standing instruction to ask before guessing, and check each part with one test.

Before you start

Menu names move. Anthropic's help pages currently point to Settings, Customize and Capabilities in different places; if a label differs on your screen, look for the nearest equivalent rather than assuming something is missing.

Step 1 — Give Claude a memory

  1. Open Settings > Memory on the web or in Claude Desktop. Check memory is on. (One help page also describes a "View and edit your memory" link under Settings > Capabilities; use whichever your screen shows.)
  2. Read what Claude has already saved. Delete anything wrong or stale.
  3. If you use another assistant, use the import flow: Settings > Memory, Start import. Run the export prompt Anthropic's help page describes in your old tool, paste the result, then click Add to memory. Anthropic describes this as experimental, and says Claude concentrates on work-related content, so personal details unrelated to work may not be kept.
  4. Add the facts you are tired of repeating by telling Claude in chat: your role, what you are working on, your tone, your tools, your spelling (British, in our case).
  5. Decide about sensitive topics. By default Claude leaves out things like health, race, religion and politics. There is an Include sensitive topics in memory setting; leave it off unless you have a reason.
  6. For anything you do not want remembered, use an incognito chat. These are not saved to history or memory on any plan.

Fill the memory with working context, not your life story.

Step 2 — Install Skills

Skills package a repeatable method, such as "how we format a client report", so Claude applies it when relevant.

  1. Make sure code execution is on: Settings > Capabilities, Code execution and file creation (Free, Pro, Max). On Team and Enterprise, an owner must enable cloud code execution and Skills under Organization settings > Plugins & skills.
  2. Go to Customize > Skills and toggle on the built-in skills you want. Anthropic lists spreadsheet, Word, PowerPoint and PDF skills. Claude applies them automatically when relevant.
  3. For your own method, package a skill as a ZIP, click + in Customize > Skills and choose Upload a skill.
  4. Only install skills from sources you trust. Anthropic's own warning: review contents first, especially code dependencies and external connections.

Start with one skill for the task you repeat most. Ten half-understood skills are worse than one you wrote yourself.

Step 3 — Connect your real data

Without a connection, Claude works from whatever you paste. With one, it can read the actual document, inbox or calendar.

  1. Open Customize > Connectors.
  2. Add one connector for the place your work lives (your drive, calendar or notes). Pick one, not six.
  3. When the sign-in screen lists permissions, read them. Limit scope where you can.
  4. Be aware Claude can read, create, change or delete data in a connected app according to the permissions you grant. Approve write actions deliberately, and when you run research modes, turn off tools that write to external apps.
  5. Custom connectors use a remote server that must be publicly reachable by Anthropic's cloud. Connect only to servers run by organisations you trust.

Step 4 — Make Claude ask first

The cheapest fix is a standing instruction. Open Settings (your initials, bottom left on the web) and find the account-wide instructions field Anthropic calls instructions for Claude. Paste a version of this, and edit it to sound like you.

You are my working assistant. My role: [YOUR_ROLE]. I write in British English with a plain, direct tone. My usual outputs: [YOUR_TYPICAL_OUTPUTS].

Before you answer any task that is more than a quick factual question:
1. Check whether you have what you need: the goal, the audience, the format, and any source material.
2. If something essential is missing, ask me up to three short questions and wait. Do not guess and carry on.
3. If you are making an assumption I have not confirmed, name it in one line at the top of your answer.
4. Where facts, prices or dates matter and you cannot verify them from my files or a source I gave you, say so rather than filling the gap.
5. Before sending, check your answer against my request and remove anything I did not ask for.

Fill in your role and typical outputs. Use project instructions as well for work specific to one project; account-wide instructions are for things true everywhere.

Check it worked

Run these four checks in a new chat.

Part Test Pass looks like
Memory "What do you know about my role and how I like answers written?" Accurate, nothing outdated
Skill Ask for the task the skill covers, without naming it Claude uses the method, output matches your format
Connector "Find [A_REAL_DOCUMENT_OR_EVENT] and summarise it" Correct item found, not a guess
Ask-first Give a vague task, such as "write the update" Claude asks questions or flags assumptions

If a test fails, fix that one part rather than adding more instructions.

What to skip

Guardrails

Sources

All 751 AI guides · JulieMango plans from £17/mo