Claude Guide

Claude is Anthropic's AI assistant. You work with Claude in plain language: writing and revising texts, analysing documents, evaluating spreadsheets, kicking off research, thinking ideas through. As of October 2026 Claude is much more than a chat window: there are Projects with stored knowledge, Artifacts as standalone results, add-ins in Excel, Word, PowerPoint and Outlook, a Chrome extension, Claude Tag in Slack and, with Cowork, the option to hand off multi-step tasks entirely. This guide walks you once through everything that matters in day-to-day work.

1.1 The surfaces

1.2 The plans

Claude comes in personal plans (Free, Pro, Max) and organisation plans (Team, Enterprise). For mid-sized companies two routes matter most: Max as a personal subscription without an admin, and Team or Enterprise with central administration, a contract and company-grade data protection. Prices change often, so they are not listed here. Check them directly with Anthropic.

1.3 Your first chat

On the left, the sidebar with “New chat”, your previous chats and the Projects. In the middle, the input field. Use the plus icon to attach files and switch on tools such as web search, Research or Connectors. Next to the send button you see the model name, which holds model, effort and thinking (chapter 2).

1. Take a real task: for example paste raw meeting notes and write above them what they should become (“decisions, open items, next steps with owners”).
2. Check the result: Are the assignments right? Is anything missing? Did Claude add something that was not in your notes?
3. Refine instead of starting over: “Shorter”, “owners in bold”, “item three is wrong, correct is …”. Claude keeps working on the same answer.
4. Check your settings: privacy toggle, instructions for Claude, memory and “Code execution and file creation”. Details follow in chapters 4, 6, 9 and 15.

Three controls decide how thoroughly, quickly and economically Claude works: model, effort and thinking. All three sit in the same menu: click the model name next to the send button. Changes apply from the next answer, even in the middle of a chat.

2.1 The current models (as of 5 October 2026)

Fable on your plan: Fable 5 and 5.1 are available on all paid plans. On Max, on premium seats in the Team plan and on premium seats in seat-based Enterprise plans they are included, and you can use up to 50 percent of your weekly limit on them. On Pro and on standard seats in the Team plan they only run via usage credits, meaning additional paid usage.

2.2 Effort: how thoroughly Claude works

2.3 Thinking: adaptive instead of a toggle

On Sonnet 5.5, Opus 5.5, Fable 5.1 and Opus 5 thinking is always on, at every effort level. Claude decides how much to think (adaptive thinking), and effort steers the depth. A separate toggle (“Extended”) only exists on older models. You see the reasoning in an expandable section above the answer, with a timer. Open it for decisions: that is where you see which assumptions Claude made.

2.4 Which setting for which task

2.5 Saving your usage limit

• Choose a smaller model and a lower effort level for routine work.
• Switch off tools and Connectors that add nothing to the current chat.
• Start a new chat for a new topic instead of continuing a very long one.
• Use Research and file creation deliberately, both consume more than a short question.
• On the Enterprise plan, admins can define since 1 July 2026 (beta) which models and effort levels you see. If a model is missing, that is usually why.

Good prompts are not magic spells. They are four habits plus one rule that applies from now on. The running example: the purchasing team lead at the (fictional) Nordlicht Logistik GmbH receives a price increase announcement from her packaging supplier.

3.1 The four habits

1. Context before instruction. First say what is going on (situation, goal, people involved, limits), then what Claude should do. “Write an email about the price increase” delivers average results, because Claude fills every gap with the average of a thousand similar emails.
2. Give examples. A real email of yours as a tone template is stronger than three paragraphs of style description. Add a counterexample: “Not like this: no superlatives, no ‘we are extremely delighted’.”
3. Break big tasks into steps. First “List the five strongest arguments, no email yet”, then sort them, then “Now the email based on arguments one and three”. One prompt, one result.
4. Refine instead of starting over. The first answer is a draft. “Shorter. Drop the second paragraph. End with two concrete date proposals.” A new chat is only worth it if the conversation has gone completely off track.

3.2 The prompt framework

For quick questions one sentence is enough. As soon as a result is reused or goes outside the company, the five building blocks pay off. You can skip the role if it is already in your instructions for Claude (chapter 4).

Role: I am the purchasing team lead at Nordlicht Logistik GmbH,
a mid-sized logistics service provider.

Context: Our packaging supplier has announced a price increase of six
percent from January. Our framework contract runs until June. We have been
a customer for eight years and are satisfied. I do not want to simply accept
the increase, but I do not want to damage the relationship. The goal of the
email is a meeting, not a negotiated result.

Task: Write an email to the head of sales asking for a meeting about the
announced increase.

Format: English, formal, friendly and direct, max. 150 words, with subject
line, ending with two date proposals for next week.
Not: superlatives, empty phrases, concrete prices as a counteroffer.

Example: This is how we sound: [paste an earlier email]

3.3 The duty to check

Everything in an answer comes from one of two sources. Either you wrote it into the prompt or provided it as a file, then it is checked. Or Claude added it because it sounds plausible, then you have to check it. Go through finished texts sentence by sentence and ask for each statement: is this in my material? Especially for numbers, names, dates, legal clauses, quotes and anything that goes outside the company.

3.4 Three checks before you send

• Does the context come before the task?
• Does Claude know what the result is for and who it is for?
• Did I say what Claude should leave out?

The biggest lever for better answers is context, and you do not have to retype it in every chat. In the field “Instructions for Claude” you store once who you are, what you do and how you want answers. Claude receives this automatically in every chat. Other tools call this custom instructions.

4.1 Where to enter them

Click your initials at the bottom left, then Settings. There you find name, occupation and the field “Instructions for Claude”. The instructions apply account-wide in all chats and Projects.

4.2 The three levels of personalisation

On top comes the individual prompt. The levels complement each other. If they contradict each other, in practice the more specific one usually wins: if your instructions say “answer briefly” and you write “please in detail” in the chat, you get detail. If Claude does something you cannot explain, go through the levels.

4.3 Template to copy

My role: [job title, department, one sentence of context]
My focus: [three to five tasks where I use Claude]
My knowledge: [where I am an expert, where I need explanations]
Language and form: [English/German, prose/bullets, formal/informal in drafts]
Answer style: [e.g. result first, then reasoning; ask when unsure]
What Claude should leave out: [e.g. no intro sentences, no summary at the end]
Typical situation: [one recurring task and how I want it handled]

I am the purchasing team lead at Nordlicht Logistik GmbH (around 320
employees). I am responsible for supplier management, framework contracts and
price negotiations. Answer in English, precisely and without filler. Result
first, then the reasoning. For decisions show me pros and cons and the risk
I am most likely to overlook. If information is missing, ask instead of
assuming it. No intro sentences, no summary at the end, no bullet lists when
prose works.

4.4 Tips

• Prohibitions are powerful: “No bullet lists when prose works” often does more than three sentences about your style.
• No customer names and no data about colleagues in the instructions. Describe your role, not your environment in detail.
• Refine after two weeks: by then you know what bothers you about the answers.
• Office add-ins have their own field: in Excel, Word and PowerPoint the add-in settings contain a separate Instructions field that only applies in that app.

A Project is a workspace for one topic. You store knowledge and rules once, and every chat in the Project has them automatically. This is the point where Claude starts to feel like a colleague who knows the job. Projects exist in all plans, up to five on the Free plan.

5.1 The two parts of a Project

The Project name and description are only for your orientation. The test question for project knowledge: would you hand this to a new colleague on day one so she can join the conversation? Then it belongs in.

5.2 Writing project instructions

This Project is about purchasing packaging material at Nordlicht.
Answer from the purchasing perspective.
For figures and terms from the project knowledge always name the source
(file and location). If information is missing, say so instead of estimating.
If documents contradict each other, give both versions with dates.
Keep answers short, result first, then reasoning. Drafts to suppliers in a
formal tone. Never phrase prices as a commitment, always as stated in the document.

Write instructions for behaviour, not for this week's task. Tasks belong in the prompt. Whatever is already in your instructions for Claude does not need repeating here.

5.3 When the Project grows

On paid plans Claude automatically switches to a search mode (RAG) when the project knowledge approaches the limit of the context window. This expands capacity by up to ten times. Claude then no longer reads everything at once but searches for the relevant passages. Three things help:

• Meaningful file names: “Framework contract Feddersen 2026” instead of “contract final final”.
• Structure: headings and clear sections are found more reliably.
• Name the document in the prompt: “Look in the Feddersen framework contract, section price adjustment”.

5.4 Working rhythm and handy moves

• Chats do not share their content automatically. Whatever lasts (a negotiation line, a decision) goes into the project knowledge as a short document.
• Tidy up: uploaded files do not update themselves. Old price list out, new one in.
• Separate memory per Project: what Claude learns in the Project stays there.
• No incognito in Projects: incognito chats only exist outside Projects. Inside a Project you can switch off memory for individual chats.
• Good scopes: one Project per product category, per major initiative or per recurring task. Not one Project called “Work” for everything.

5.5 Sharing Projects in a team

On Team and Enterprise you share Projects with individuals, via an email list or with the whole organisation. “Can view” means: see the contents and chat in the Project, but change nothing. “Can edit” means: manage instructions, knowledge and access. Admins can switch sharing off entirely, existing shares then remain active. On the Max plan there is no sharing, your Project belongs to you.

When the result is more than an answer you read and forget, Claude puts it into its own window next to the chat as an Artifact: a document, a table, a chart, a dashboard or a small tool such as a calculator. The conversation stays on the left, the work piece sits on the right.

6.1 How to work with Artifacts

• Trigger: Claude creates Artifacts by itself when the content is standalone and longer. It is safer to add “as an Artifact”.
• Change: write in the chat below what should be different. Claude updates the Artifact, the version picker takes you back.
• Find again: all Artifacts land automatically in the “Artifacts” area in the sidebar.
• Template types: on paid plans there are Docs (documents you edit together), Slides (exportable as PowerPoint or PDF) and Design (visuals, mockups, one-pagers in your own design system).

6.2 Ten everyday ideas

6.3 Publish and share

6.4 Creating finished files

With the same toggle Claude creates real files to download: Excel (.xlsx) with working formulas, Word (.docx), PowerPoint (.pptx) and PDF. Maximum 30 MB per file. Example: “Create an Excel file from the order history with one sheet per supplier and totals per quarter, plus a Word document with a half-page summary for management.” On Team and Enterprise, Owners decide whether the code environment may access the internet: not at all, package sources only (default) or additionally approved domains.

You add files via the plus icon in the input field, by drag and drop or, for screenshots, straight from the clipboard. A file belongs to the chat you upload it to. If you need it again and again, it belongs in the project knowledge.

7.1 Formats and limits

The real bottleneck is rarely file size but the amount of content Claude can keep in view at once. A 400-page contract is harder work than a large photo. Ask specifically (“only chapters 4 to 6”) or put large amounts of material in a Project.

7.2 Use cases by file type

• PDF: check offers for terms (“prices per tier, deadlines, changes from last year, with page numbers”), compare two contracts, scan tender documents for mandatory items.
• Excel and CSV: analyse order history by supplier and quarter, find and explain outliers, standardise spellings, build charts.
• Word and text: summarise draft contracts, revise your own texts, modernise old templates.
• Images and screenshots: read type plates and delivery notes, describe transport damage, turn a screenshot of a price list into a table, transcribe handwriting.
• Transcripts: summarise conversations by agreements, commitments on both sides, open items and owners.

7.3 Before you upload

• Everything in the file goes to Claude, including what you do not need. Would an excerpt do?
• Does it contain personal data that adds nothing to the task (names, phone numbers, signatures)?
• Whatever Claude extracts from a file, you check against the file. So always ask for page references.

Without web search Claude answers from its own knowledge, which has a cutoff date. For anything current you need one of three sources: the open web, Research for big questions, or your own documents via Connectors and Enterprise Search.

8.1 Web search

Switch it on via the plus icon, menu item web search. Claude then searches when needed and adds citations to its statements. For timeless questions Claude answers directly; if you want to be sure, add “search the web”. Good for a specific figure, a news item, a term. Duration: seconds.

8.2 Research

• Plans: Pro, Max, Team and Enterprise, on the web, in Claude Desktop and on mobile.
• Prerequisite: web search must be switched on.
• Switch on: plus icon, “Research”. A blue indicator shows that Research is active.
• How it works: Claude searches many times in a row, each search building on the previous one, and delivers a report with checkable citations. Duration: minutes.
• Sources: the web plus connected services such as Gmail, Google Calendar and Google Docs.
• Usage: Research counts like a normal chat but uses your limit faster.
• If Claude does not research: write “Use the research tool to …” in the prompt.

8.3 Enterprise Search

On Team and Enterprise, Enterprise Search provides a preconfigured Project “Ask Your Org” in the sidebar that searches across your tools, such as Slack, Microsoft 365 and custom Connectors. An Owner sets it up once and selects Connectors for documents and chat, email is optional. Then everyone signs in to the services they want to search. Anthropic maintains the instructions behind it.

8.4 Which source when

8.5 Checking sources in under a minute

1. Click. A source you have not opened is not a checked source.
2. Date. Does the date fit your question?
3. Publisher. Statistics office, association, trade media or a vendor with its own interest?
4. Statement. Does it really say that? “Up to” easily becomes “on average”, “planned” becomes “decided”.
5. Origin. Primary source or copy? Three hits with the same original source are one source.
6. Gap. Is the opposing view missing? Are there figures for your country or only for the US?

Three features determine what Claude knows about your earlier conversations. You find all three under Settings > Memory.

9.1 Memory

Claude remembers things from your conversations, such as what you are working on or how you like answers, and carries them into later chats. Since 10 July 2026 memory consists of individual entries sorted by topic instead of daily summaries. Under “Topics” you open each entry, edit it or delete it. In the chat, “Remember this” or “Forget that again” also works.

• “Generate memory from chats”: Claude creates memory from your chats.
• “Include sensitive topics in memory”: by default Claude leaves out topics such as health, ethnicity, politics or gender identity. If you switch them on, you see a notice for each entry saved on them.
• Memory off for one chat: before the first message via the plus menu.
• Projects: each Project has its own memory, separate from your other chats.

9.2 Chat search

With “Search and reference chats” Claude searches your earlier conversations on request (“What did we discuss last week about the price adjustment clause?”). Available on Pro, Max, Team and Enterprise, on the web, in Claude Desktop and on mobile.

9.3 Incognito chats

• Start: ghost icon at the top right of a new chat outside a Project. A black border shows the mode.
• Not in history, not in memory, not in chat search and not in the monthly recap.
• No training, even if you have switched on model improvement in the privacy settings.
• Retention: 30 days, on the Enterprise plan according to your organisation's policy.
• Final: a closed incognito chat cannot be reopened.
• Available on all plans, but not inside Projects.

With the Office add-ins Claude comes to where you already work. The difference from the chat window: you upload nothing, Claude sees the open file and works directly in it.

10.1 Status and plans

All four are available for Pro, Max, Team and Enterprise. They run in the web version and on Windows and Mac with a Microsoft 365 subscription, not in Office 2016/2019 and not on iPad or Android.

10.2 Installing

• Yourself: open the listing “Claude for Microsoft 365” in Microsoft AppSource (Excel, Word, PowerPoint), “Get it now”, then activate the add-in in the app and sign in with your Claude account. Outlook has its own listing “Claude for Outlook”.
• For the organisation: admins deploy the add-ins via the Microsoft 365 Admin Center (Integrated apps) or via a manifest file.
• Outlook in addition: an Entra Global Administrator grants Microsoft Graph consent once. Otherwise everyone sees “Need admin approval”.

10.3 Working across files

With the setting “Let Claude work across files” Claude carries the conversation across Excel, Word, PowerPoint and Outlook. In PowerPoint you say “Take the figures from the open Excel file” without copying anything.

• On by default in Pro and Max, off by default in Team and Enterprise. Owners control it under Organization settings > Office agents (“Let Claude work across apps”).
• The setting is per device, so check it again on a second computer.
• Claude only reaches files that are currently open and cannot open, close or switch files itself.
• Skills you have enabled also work in the add-ins.

10.4 What you need to know

• Separate instructions per app: each add-in has its own Instructions field in its settings.
• Outlook never sends by itself: every draft lands unsent in the compose window, you click Send.
• Chat history stored locally: the history sits in browser storage, not on Anthropic's servers, and is not synced between devices.
• Data: inputs and outputs are deleted on the backend within 30 days. Custom retention periods of the organisation do not apply to the add-ins, and their activity is not in the Enterprise audit logs.
• Prompt injection: external files and emails can contain hidden instructions. Only work with trusted files and read every confirmation prompt carefully.

11.1 Claude in Chrome

Claude in Chrome is an extension from the Chrome Web Store. Claude works in a side panel next to the web page you are on: summarise product pages, compare suppliers, understand portals and forms, without copying.

• Plans: all paid plans (Pro, Max, Team, Enterprise).
• Side panel as a Cowork session: on Max and Team the panel runs as a Cowork session, on Pro this is rolling out. On Enterprise only once an admin has enabled Cowork in the cloud, until then the classic view applies.
• Caution: Claude sees what is on the page. Treat sensitive pages as deliberately as an upload. What is on a web page is a source like any other.

11.2 Chrome controls for admins

Under Organization settings > Claude in Chrome, admins define an allowlist (only these sites) and a blocklist (never these sites). Anthropic recommends starting with a tight allowlist and expanding it step by step.

11.3 Claude Tag in Slack

Since 23 June 2026, Team and Enterprise organisations can tag Claude directly in Slack conversations and delegate tasks. Claude Tag is currently in beta.

• @Claude in a thread or channel: Claude reads the conversation context and completes the task, such as turning a discussion into action items or running a calculation.
• Standing instructions: Claude can also work without a mention, such as watching channels, posting weekly digests or flagging urgent items.
• Own identity: Claude acts in your systems under its own account. Admins define once which tools Claude may access in which channel.
• Microsoft Teams: announced, currently waitlist only.

Connectors link Claude to tools you already use. Claude can search, read there and, depending on permissions, also act. This chapter is the overview; setup and permissions in detail are in the separate Connectors guide in the toolkit.

12.1 Three types

12.2 Switching on

• As a user: under Customize > Connectors or in the chat via the plus icon, “Manage connectors”.
• On Team and Enterprise: an Owner activates the Connector under Organization settings > Connectors. Then everyone signs in with their own account.
• Permissions per tool: “Always allow”, “Needs approval” or “Blocked”. That way Claude may, for example, search emails but not send them.
• Claude never sees more than you: what you are not allowed to open in the source system, the Connector cannot reach either.

12.3 Microsoft 365

The Microsoft 365 connector links SharePoint, OneDrive, Outlook and Teams, on all plans. By default it is read-only. Write tools (sending emails, managing drafts and events, creating files in OneDrive and SharePoint) need separate approval from the admins. Before first use, an Entra Global Administrator must approve the integration once for the tenant.

A Skill is a stored guide for a recurring task: for example customer emails in house style, a meeting note in a fixed format or the weekly purchasing report. Claude only reads the name and description of all Skills and loads the rest when a request fits. The detailed guide with example Skills is on the separate Skills page in the toolkit.

13.1 Prerequisites

• Skills are available on all plans. The prerequisite is the toggle “Code execution and file creation”.
• Switch them on under Customize > Skills, each Skill individually.
• On Team and Enterprise an Owner checks that “Code execution and file creation” and “Skills” are active under Organization settings > Plugins & skills.
• Anthropic's Skills for Excel, Word, PowerPoint and PDF are included. Claude uses them automatically when you create files.

13.2 Your own Skills

• Structure: a folder with a file SKILL.md, optionally with supporting files such as templates, references or scripts.
• name: at most 64 characters, only lowercase letters, digits and hyphens. The words “claude” and “anthropic” are not allowed.
• description: at most 1,024 characters and must say what the Skill does and when it applies. It decides whether Claude loads the Skill.
• Upload: add the Skill folder as a ZIP via Customize > Skills.
• Share: on Team and Enterprise share with individuals or groups, or publish to the organisation's library.

---
name: supplier-meeting-notes
description: Turns notes or transcripts of a supplier meeting into meeting
  notes with agreements, commitments on both sides, open items and owners.
  Use when meeting notes, minutes or a summary of a supplier conversation
  are needed.
---

# Supplier meeting notes

## Steps
1. Name participants, date and occasion at the top.
2. Four sections: Agreed, Supplier commitments, Our commitments, Open.
3. Add the name of the responsible person to each item.
4. Add nothing that is not in the material. Mark gaps as open.

With Cowork you hand off multi-step tasks: Claude plans, works through the steps itself and delivers finished results, such as Excel files with working formulas, PowerPoint presentations, formatted documents or prepared research. This chapter is the overview; details are in the separate Cowork area of the toolkit.

14.1 How Cowork works

• In the cloud: Cowork runs in an isolated environment on Anthropic's servers. Work continues when you close your laptop and follows your account across devices.
• Desktop only for local work: Claude Desktop must be open when Claude accesses connected local folders, local Connectors, the browser or computer use.
• Approval modes for actions: Manual (Claude asks for every action), Auto (Claude reviews each action for safety and then proceeds) or Skip (no checks and no approvals).

14.2 Chat and Cowork are one

Since 16 September 2026 a shared interface has been rolling out gradually on Pro and Max: there is no longer a separate choice between chat and Cowork. You say what you need, and Claude decides whether that is a quick answer or a task. On the web and on mobile, Cowork is in beta for Pro, Max and Team, on Enterprise only if an admin has enabled it.

14.3 Scheduled tasks

On all paid plans you can run tasks on a recurring basis: “Scheduled” in the sidebar, then “New task” with Claude or set it up manually. Frequency hourly, daily, weekly, on weekdays or manual. The tasks run on Anthropic's side, your computer may sleep.

The most important question for mid-sized companies is not what Claude can do, but under which conditions your data is processed. Here personal plans and organisation plans differ fundamentally.

15.1 Max (and Free, Pro): you decide yourself

• Model improvement toggle in the privacy settings. If it is off, new chats are not used for training. Deleted chats disappear from history immediately and from the backend systems within 30 days.
• If it is on, new chats may be used for training and kept in de-identified form for up to five years in training pipelines.
• For professional use: toggle off, and check this right after your first login.

15.2 Team and Enterprise: contract and central control

• No training: inputs and outputs from commercial products are not used for training by default. Exception: explicit feedback via thumbs up or down.
• DPA: the data processing addendum with standard contractual clauses is part of the Commercial Terms. Your company is the controller, Anthropic the processor.
• Retention on the Enterprise plan: without a custom setting, data is kept indefinitely. Primary Owners or Owners can set custom periods from 30 days under Organization settings > Data and Privacy.
• Memory is off by default, Owners enable Connectors and Skills.

15.3 Side by side

15.4 Admin tools on Team and Enterprise

• Organization settings: Capabilities, Connectors, Plugins & skills, Office agents, Claude in Chrome, Data and Privacy.
• Roles and permissions: role-based access for Enterprise (since 9 April 2026), admin permissions in custom roles (2 June 2026), Connector permissions in custom roles (28 May 2026).
• Model entitlements: since 1 July 2026 (beta) Enterprise admins control which models and effort levels users see.
• Plugins: bundle Skills, Connectors and further building blocks. Owners distribute them with the options “Installed by default”, “Available to install”, “Not available” or “Required”.
• Review and evidence: Compliance API, security scanning for Skills and plugins (Enterprise, since 6 August 2026), smart reports on usage and costs (Enterprise, beta since 10 September 2026).

16.1 Glossary

16.2 The key releases of 2026

16.3 Keep learning

• Claude Help Center: support.claude.com, with release notes for every new feature.
• Claude Academy: academy.claude.com, free courses and tutorials from Anthropic's education team.
• Privacy: privacy.claude.com for retention, training and the DPA.