Learning

Understanding Anthropic, feature by feature.

A practical introduction to what Anthropic offers, how to instruct its assistant, and how each capability fits into everyday work.

What Anthropic actually offers

An assistant you instruct in plain language.

Understands language

Reads and produces written work to a professional standard. You instruct it in ordinary language — no specialist syntax is required.

Reasons through problems

Works through multi-step problems rather than returning a single undifferentiated response.

Works with local files

Documents can be provided directly, and questions asked about their contents.

Acts through tools

Subject to approval, connects to the systems and applications already in use across the organisation.

Anthropic's assistant produces drafts. Judgement and final decisions should always remain with the person using it.

Getting the instruction right

The quality of the instruction determines the quality of the result.

1

Give context

Explain the background, state the purpose and identify the reader.

2

Be specific

Specify the format and length you want. Vague instructions produce non-specific results.

3

Show an example

Provide a sample or pattern to match. Demonstrating the requirement beats describing it.

4

Refine the output

Treat the first response as a draft. Identify what is lacking and request a revision.

Instead of this

Try this

Write a job ad

Draft a 250-word ad for a part-time warehouse assistant, friendly tone, three bullet requirements.

Summarise this

Summarise this in five bullets for a finance director with two minutes.

Make it better

Tighten this to under 150 words and remove the jargon.

Chats and Projects

Begin in a chat. Move to a Project once the task becomes recurrent.

As a general principle: if the same background would be explained twice, it belongs in a Project.

A chat

  • Suited to one-off questions and short drafts
  • Requires no configuration
  • Context is retained for that conversation only
  • Appropriate while requirements are still being established

A Project

  • Holds instructions and reference material that apply to every chat within it
  • Suited to recurring work: a weekly report, an established tone of voice, a defined process
  • On Team and Enterprise plans, Projects can be shared across the organisation

Instructions

Standing guidance on tone, format and rules, applied to every chat within the Project.

Knowledge

Reference material to draw upon: brand guidelines, process documentation, worked examples.

Continuity

Context from earlier work is carried forward, removing the need to restate background.

The Anthropic toolkit

One assistant, accessed through several interfaces.

Select the one closest to where the work resides.

Chat and Projects

The everyday workspace for drafting, analysis, enquiries and reusable workspaces.

Research

For enquiries requiring more than a single search. It plans the investigation, conducts it and reports findings with cited sources. Useful for market scans, supplier comparisons and background briefings.

Connectors

Allow access to systems already in use, such as Google Drive, SharePoint and Slack. New connectors are subject to security review and scoped to the minimum access required.

Claude Tag

The assistant works inside Slack under its own identity. A team can hand it a task by typing @Claude in a channel, with the exchange visible to everyone there. Available on Team and Enterprise plans.

Cowork

Designed for multi-step knowledge work — research, analysis and document production — rather than software development. It can access live files in connected folders, so only connect what the task genuinely requires.

Claude Code

A coding agent available in the terminal, within an IDE, in the desktop application and on the web. It reads the repository, works across multiple files, stages changes and reviews diffs.

Claude in Chrome

A browser extension that reads and acts upon web pages alongside the user. It navigates, selects, enters text and completes forms. Because it reads third-party content, treat its output as information, never as instruction.

Excel and Microsoft 365

The assistant operates within the documents themselves, via a sidebar rather than a separate application, so work stays where the data already lives.

Artifacts

A document, interactive view or small tool that the assistant builds and refines inside the conversation, rather than as plain text in the reply. Shared artifacts should be checked before being relied upon.

Skills

A set of instructions and reference material the assistant loads only when it is relevant to the task at hand. They keep responses focused and avoid repeating setup instructions.

Plugins

A bundle of Skills, connectors and tools pre-configured to support a single business function, such as sales, finance, legal, marketing, HR or engineering.

Choosing the right one

An indicative guide.

If you want to…
Use this
Draft, summarise or work through a problem
Chat
Investigate across several sources with citations
Research
Recurring work that needs shared instructions
Projects
Hand a task to an assistant inside Slack
Claude Tag
Multi-step document or analysis work across files
Cowork
Modify code, or work within a spreadsheet
Claude Code, or Claude for Excel
Build a document, visual or small tool without leaving the chat
Artifacts

Where the choice remains unclear, begin in a chat: it is the least costly way to establish what is required.

Using Anthropic safely

Five practices that address the majority of the risk.

Do this
Not this
Verify facts and check sources
×Treat output as automatically accurate
Use organisational accounts for company work
×Use personal accounts for company work
Confirm outputs before acting on them
×Paste confidential or unapproved data without review
Escalate misuse or uncertainty
×Assume accountability lies with the tool
Seek guidance where uncertain
×Rely on generated output without judgement

Accountability remains with the user. That output was AI-generated is not a defence.

Capabilities are set at organisation level. Scheduled and cloud-executed tasks are individually approved and assigned a named owner. Where a required feature is unavailable, a request route should exist through your internal owner.

Getting started

Your first week.

1

Select one weekly task and attempt it in a chat.

2

Provide context, be specific, and refine twice before forming a judgement.

3

Once the approach works, establish it as a Project.

4

Share what you learn with one colleague so good practice spreads.