Claude Projects, Skills, Code and Cowork compared

Claude gives a business five working surfaces, and the practical question is allocation, not ranking. Chat with Projects is the default for everyday knowledge work; Skills capture procedures a team repeats; Claude Code belongs to engineers and code-shaped work; Cowork delegates multi-step tasks to an agent that keeps working when you step away; and Claude for Microsoft 365 puts Claude inside Excel, PowerPoint, Word and Outlook. All five run the same underlying models. What differs is the shape of work each is built for, which plans include it, and what each asks of governance. Deciding which staff get which surface, in which order, is most of what a good rollout amounts to.

This guide is for the operations lead or internal champion making that allocation. Product facts are taken from Anthropic's current documentation, accessed July 2026 and listed under Sources; several of these surfaces are moving quickly, so recheck availability notes before committing a rollout plan to paper.

Which Claude surface should a business use for which job?

Most staff need only the default surface, a few need one specialist surface, and almost nobody needs all five. Allocating deliberately also keeps training focused, because each surface asks users to learn a slightly different way of working. The table below is the allocation summary; the sections that follow take each surface in turn.

Surface	Built for	Best-fit jobs	Plan availability	Governance notes
Chat with Projects	Everyday assisted work with per-workflow context	Drafting, analysis, research, document questions	Projects on all plans (five on Free); sharing on Team and Enterprise	Approved documents only in knowledge bases; organisation instructions apply everywhere
Skills	Repeatable procedures and house formats	Meeting-note formats, proposal structures, brand and style checks	All plans; requires code execution enabled	Owners can provision organisation-wide; peer sharing is off by default
Claude Code	Engineers in repositories and terminals	Coding, data pipelines, repository automation	Included in all paid plans	An agent that edits files and runs commands, so scope it with engineering-grade permissions; not a general-staff surface
Claude Cowork	Delegated multi-step knowledge work	File organisation, research synthesis, spreadsheets and decks from messy inputs	Paid plans; desktop generally available, web and mobile in beta	Three permission modes; heavier usage; activity not yet in the Compliance API
Claude for Microsoft 365	Staff whose day happens in Office	Excel models, decks in the house template, tracked-change edits, inbox triage	Word, Excel and PowerPoint on all paid plans; Outlook in beta	Reviewable by design: tracked changes, highlighted cells, drafts that wait for send

When is Chat with Projects enough?

For most everyday knowledge work, most of the time. A project gives each workflow its own workspace: instructions tailored to the job, a knowledge base of approved documents, and a chat history that stays together, so the context does not have to be re-explained every morning. That combination covers the bulk of what a typical UK SME actually does with Claude in its first year: drafting, summarising, analysing and asking questions of its own material. On Team and Enterprise plans, shared projects let a whole team work inside the same context with viewer and editor permissions kept separate.

The rollout implication is simple: start everyone here, one project per distinct workflow, and escalate individuals to specialist surfaces only when a specific shape of work demands it. The projects themselves generate the evidence for that escalation: a project whose chats keep repeating the same formatting instructions is asking to become a skill, and a project that keeps receiving the same multi-step request every Monday is describing a Cowork task. Reviewing surface allocation against real usage after the first quarter costs an hour and typically moves a few people onto a better-fitting surface. Firms whose main ambition is interrogating their own document estate can get a long way without ever leaving this surface; our guide to Claude for internal knowledge management covers that pattern in depth.

When should a team reach for Skills?

When the same procedure or format is being pasted into chats for the third time. A skill packages a repeatable method, the house way of structuring meeting notes, qualifying a tender, or formatting a client report, so Claude applies it consistently without anyone re-typing instructions. Anthropic's documentation draws the dividing line against projects cleanly: a project supplies standing background knowledge inside its own workspace, while a skill supplies a procedure that loads when a task needs it and works everywhere. Skills are available on every plan and require code execution to be enabled.

The governance shape is what makes skills interesting for firms rather than hobbyists. On Team and Enterprise plans an owner can provision an approved skill to every member centrally, while peer-to-peer skill sharing is switched off by default, which means a business can run skills the way it runs document templates: one approved version, distributed from the centre, rather than forty private variants. The fuller question of when a workflow deserves a skill, and when it does not, is answered in when to build a Claude Skill, and writing better instructions in general is covered in our prompt engineering guide for UK business teams.

Who should use Claude Code, and who should not?

Engineers, and operators whose work is genuinely code-shaped; nobody else. Claude Code brings the same models to repositories, terminals and data pipelines, delegating coding tasks from concept to completion, and it is included in every paid Claude plan, so there is no licensing reason to keep it from a technical team. There are, however, governance reasons to scope it deliberately: a coding agent acts on files and runs commands, so it needs the permission boundaries and review habits that engineering teams already understand. Our coding agent governance guide sets out that control set, and for teams comparing tools, OpenAI Codex vs Claude Code covers the head-to-head.

The test for the grey area, technically comfortable staff who are not developers, is whether the work involves structured files and repeatable transformations. If it does, Claude Code can suit them; if the work is documents, meetings and email, the surfaces on either side of this section will serve them better with less to govern.

What is Claude Cowork actually for?

Delegation. Cowork takes the agentic architecture that powers Claude Code and applies it to knowledge work without a terminal: you describe an outcome, step away, and come back to finished work such as organised files, synthesised research, spreadsheets with working formulas, or a first-draft deck. Sessions run remotely, a capability Anthropic currently labels beta, in an isolated environment on Anthropic's servers, and keep working when the laptop closes; on desktop, Claude can read and write local files you have granted it; complex jobs are broken into parallel workstreams; and tasks can be scheduled to run on a cadence with no device online. You stay in the loop while it works: progress and reasoning are visible, you can steer mid-task from any device, and permanent file deletion always requires an explicit approval. Cowork also carries its own instruction layer, with global instructions applying to every session and folder-level instructions adding job-specific context, which slots neatly into the layered setup a well-run installation already has (the layers themselves are mapped in our guide to setting up Claude for a business). It is available on paid plans, generally available on desktop, with web and mobile in beta.

The governance profile is different from chat, and worth stating plainly. Cowork has three permission modes: manual approval of each action, an automatic mode in which Claude reviews its own actions for safety and blocks what fails the check, and a skip mode with no checking that Anthropic itself advises reserving for fully trusted work. On Team and Enterprise plans an organisation can require per-task approval for connector tools that write to live systems, and admins can monitor Cowork activity across the organisation via OpenTelemetry; Anthropic's documentation also notes that Cowork activity is not captured in the Compliance API at this time, which compliance-led firms should factor into their audit posture. Delegated tasks also consume noticeably more usage than chat. The practical read: Cowork rewards teams whose workflows and boundaries are already defined, and it should be given structured delegation with clear inputs, not vague hopes. The National Cyber Security Centre's AI guidance points the same way: security with agentic tools is as much about organisational process as about the technology.

Where does Claude for Microsoft 365 fit?

With the people whose entire day happens inside Excel, PowerPoint, Word and Outlook, which in many UK firms is most of the office. Claude works inside the four applications: building and editing Excel models, drafting slides in the house template, editing Word documents, and triaging an Outlook inbox. Word, Excel and PowerPoint support is generally available on all paid Claude plans, with Outlook in beta, and firms with cloud commitments can connect through Microsoft Foundry, Amazon Bedrock or Google Cloud Vertex AI. The distinctive capability is continuity: one conversation carries context across the four applications, so a request that starts in the inbox can open the brief in Word, build the model in Excel and produce the deck in PowerPoint without the context being re-explained at each step.

Its most under-appreciated property is that it is reviewable by design: Word edits arrive as tracked changes, Excel changes are highlighted, and Outlook drafts open in the native compose window and wait for a human to press send. For a firm writing human-in-the-loop rules into its AI policy, this surface has the control built in rather than bolted on. Skills also carry across the four applications, so a house format defined once applies whether the output is a document, a deck or a spreadsheet.

How should a business choose in practice?

Start everyone on Chat with Projects, then move specific people to specific surfaces as the work demands it: a procedure repeated weekly becomes a provisioned skill; repository work goes to Claude Code under engineering governance; a recurring, well-defined delegation becomes a Cowork task once the workflow is understood; and Office-centric staff get Claude for Microsoft 365. Every surface sits under the same governance frame, the data classification, approved documents and named sign-off that apply to the whole installation, so adding a surface never means renegotiating the rules from scratch. Note what the allocation is not about: budget. Claude Code, Cowork, skills and the Microsoft 365 applications are all included in the paid plans rather than sold separately, so the constraint is governance capacity and training attention, not licence count. That is an argument for sequencing surfaces deliberately rather than switching everything on at once.

The AI Consultancy is an Anthropic Consulting Partner and makes exactly this allocation, surface by surface and role by role, as part of Claude deployments for UK organisations. If you want the map drawn against your actual workflows rather than a feature list, our Claude consulting and Claude implementation services do that work, and the Knowledge Hub training section collects the supporting guides.

Sources

• Claude Help Center, "Get started with Claude Cowork", accessed July 2026 (agentic architecture without a terminal, paid-plan availability, desktop general availability with web and mobile in beta, remote isolated sessions, sub-agent coordination, local file access, scheduled tasks, the three permission modes, per-task approval for write-capable connector tools on Team and Enterprise, OpenTelemetry monitoring, Compliance API exclusion at this time, higher usage consumption).
• Claude product page, "Claude for Microsoft 365", accessed July 2026 (Excel, PowerPoint and Word generally available on all paid plans with Outlook in beta, tracked changes and highlighted cells, drafts that wait for send, context across applications, template preservation, deployment via Microsoft Foundry, Amazon Bedrock and Google Cloud Vertex AI, skills across the four applications).
• Claude Help Center, "What are skills?", accessed July 2026 (dynamic loading, availability on all plans, code execution requirement, organisation-provisioned skills, the projects-versus-skills distinction).
• Claude Help Center, "Provision and manage skills for your organization", accessed July 2026 (owner provisioning, peer-to-peer sharing off by default).
• Claude Help Center, "What are projects?", accessed July 2026 (per-project instructions and knowledge bases, five-project cap on Free, sharing and permission levels on Team and Enterprise).
• Claude Help Center, "What is the Team plan?", accessed July 2026 (Claude Code included, delegating coding tasks from the terminal).
• National Cyber Security Centre, "AI and cyber security: what you need to know", accessed July 2026.