Claude Skills are written procedures Claude loads whenever a matching task comes up, so it does the job the same way every time. Projects hold the files and instructions for one body of work, Connectors let Claude read and act in apps such as Google Drive, Microsoft 365 or Slack, and Cowork is Claude carrying out a multi-step task itself. For an operations team, the order is: write the procedure, turn it into a Skill, and keep a person approving anything that changes a system of record.
This guide describes the products as of November 2026, from Anthropic’s help pages. They change often, so check the current plan page before you rely on any feature.
The four pieces in plain language
| Feature | What an operations team uses it for | Where it stops |
|---|---|---|
| Skills | A procedure followed the same way every time: checks, reports, standard replies | Only as good as the procedure written in it |
| Projects | Context for one area of work: a supplier, a month-end, a customer account | Its files and instructions stay inside that Project |
| Connectors | Reading the shared drive, the inbox, the calendar or the team channel | Only the actions the connector offers, within your permissions |
| Cowork | Jobs with several steps: gather files, compare, build the spreadsheet, draft | Your own folders, browser and apps only while the desktop app is open |
Skills
Claude comes with built-in Skills for Excel, Word, PowerPoint and PDF files, and you can add your own. Each Skill starts with a name (up to 64 characters) and a description (up to 200 characters) that Claude uses to decide when the Skill applies: the trigger in an SOP. According to the help centre, Skills are available on the Free, Pro, Max, Ultra, Team and Enterprise plans and need code execution switched on. You add them under Customize > Skills, by uploading a ZIP of the Skill folder or by asking Claude to build one from your description.
On Team and Enterprise plans you can share a Skill with colleagues (view only) or publish it to the whole organisation.
Projects
A Project keeps the files, instructions and chats for one piece of work together. Every plan has Projects (free accounts up to five), and on Team and Enterprise plans they can be shared with view or edit rights. Anthropic draws the line clearly: a Project’s context exists only inside that Project, while a Skill is available in any conversation. Use a Project for what you are working on and a Skill for how the work is done.
Connectors
Connectors let Claude access your apps, retrieve data and take actions, such as searching Google Drive or posting in Slack. Claude inherits each person’s permissions from the connected service. For each connector, tools are grouped into read-only and write or delete, and each can be set to Always allow, Needs approval or Blocked. On Team and Enterprise plans an owner sets this for everyone, which means you can let Claude read a system while stopping it from writing anything back.
You can also add a custom connector that points to a remote MCP server. Claude connects to it from Anthropic’s cloud, so it must be reachable over the public internet: a system behind your firewall or VPN will not connect unless IT opens access.
Cowork
Cowork brings the agent behind Claude Code to everyday work, without a terminal: you describe the task, review Claude’s approach and let it run. It is available on the paid plans (Pro, Max, Ultra, Team and Enterprise). The work runs on Anthropic’s servers in an isolated environment and keeps going if you close your laptop. To work with folders on your own computer, use your browser or operate other apps, the Claude desktop app for macOS or Windows must be open. You can also type /schedule in a task to make it recurring, and scheduled tasks run in the cloud with no device online.
Cowork has three permission modes: Manual (Claude asks you to approve actions), Auto (Claude checks each action for safety itself) and Skip (no checks, only for tasks you fully trust). In every mode it asks before permanently deleting files. Anthropic is also merging chat and Cowork into one conversation (rolling out on Pro, Max and Ultra, and on Team and Enterprise where an admin turns it on), so you may not see a separate Cowork option at all.
How they fit together for repeatable work
Take a weekly late-deliveries report. Today someone exports open purchase orders from NetSuite, compares them with the carrier spreadsheet and writes a summary. Split up:
- The Skill holds the procedure: what counts as late, how to group the results, and what to do when a date is missing.
- The Project holds the context: the supplier list, last month’s reports, the agreed delivery terms.
- The Connector reaches the shared drive where the weekly export lands, read-only.
- Cowork runs it every Monday morning and leaves a draft report and a list of questions for a person to check.
The Skill makes the work repeatable. Without it, the result depends on who wrote the prompt that week.
A Skill is an SOP that Claude reads
If your team already has a written procedure, most of the Skill is done. The sections of a good SOP (see how to write an SOP people follow) map almost one to one onto a Skill.
| SOP section | Skill section | What changes |
|---|---|---|
| Purpose and trigger | Description | Written in the words people use when they ask for the job |
| Scope | When to use, when not to use | Say plainly what Claude must refuse or hand back |
| Systems and records | Inputs | Name the exact files, columns and fields it will receive |
| Steps | Steps | One action each, in the imperative, as in the SOP |
| Exceptions | Rules and exceptions | Each one ends in an action: fix it, or stop and flag it |
| Roles | What to hand back | What the person approving receives, in what format |
| Owner, version, review date | Same | Kept at the bottom of the Skill, updated when the SOP changes |
Two rules from mapping processes carry over. First, no quote, no claim: every rule in the Skill should come from someone who does the job, not from the ERP manual. Second, write the procedure people actually follow on a Tuesday when a supplier is late, not the one in the quality binder. A Skill built from the official version will be confidently wrong in exactly the cases that cost money.
Claude follows the happy path well. It goes wrong where people use judgement nobody wrote down: the supplier who always invoices freight separately, the customer whose PO numbers never match. List those cases first (our lesson on documenting the exceptions shows how), and for any case not on the list, tell Claude to stop and ask rather than guess.
How to turn an SOP into a Skill
- Pick one SOP that has been tested on a new starter. If a person could not follow it without asking, Claude will not either.
- Write the description as the trigger. Use the phrases people type: “check this invoice”, “match these invoices to the POs”. If Claude does not pick the Skill up when expected, the description is usually the problem.
- Copy the steps and exceptions across using the table above, with the column and field names in your exports.
- Add one worked example and define what comes back: a file, a table, and a separate list of anything a person needs to decide.
- Test it on three past cases, including one that went wrong, and compare Claude’s result with what your team did.
- Install it under Customize > Skills: the ZIP must contain the Skill folder at its root, and the folder name must match the Skill’s name. Or paste the SOP into a conversation and ask Claude to build the Skill from it.
- Give it an owner and a review date. When the SOP changes, the Skill changes the same day.
Example Skill: check supplier invoices against purchase orders
Replace everything in brackets. The matching logic follows our three-way match guide.
---
name: check-supplier-invoices
description: Check supplier invoices against the purchase order and goods receipt before approval. Use when someone asks to check, match or review supplier invoices.
---
When to use
When given one or more supplier invoices (PDF) with the purchase order export and the goods receipt export for the same period.
When not to use
Services and subscriptions with no goods receipt [handled by two-way match]. Credit notes. Never approve, post or pay anything: this Skill only prepares the check.
Inputs
Invoices (PDF). PO export [file name, columns]. Goods receipt export [file name, columns].
Steps
- Read each invoice and extract supplier, invoice number, date, PO number and lines.
- Find the PO by PO number. No PO number on the invoice: exception A.
- Match each invoice line to the PO line by item code.
- Compare invoiced quantity with quantity received, not quantity ordered.
- Compare unit price with the PO price, using the tolerance below.
- Check the invoice number has not already appeared for this supplier.
- Record the result for each invoice: matched, or the reason it is not.
Rules
Price tolerance: [e.g. 2% or 10 per line, whichever is lower]. Quantity tolerance: [e.g. none]. Freight lines: [e.g. accepted only if on the PO]. Never change a figure in the source files.
Exceptions
A. No PO number: list it for [AP clerk] to identify.
B. Quantity received lower than invoiced: flag for [buyer].
C. Price outside tolerance: flag for [buyer] with both prices.
D. Possible duplicate: flag for [AP lead]; do not mark as matched.
E. Anything not covered here: stop on that invoice and ask. Do not guess.
What to hand back
1. A spreadsheet with one row per invoice: supplier, invoice number, PO number, total, result, exception letter, note.
2. A short list of the invoices that need a person, grouped by who must decide.
Owner and version
Owner: [AP lead] · Source SOP: [AP-004 v2.1] · Version: [1.0, date] · Next review: [date]
What to keep a human on
A Skill can prepare the work. Some decisions should stay with a named person however good the output looks:
- Approving payments, credit notes and anything that moves money.
- Changing master data, above all supplier bank details.
- Anything sent to a customer or supplier in your company’s name.
- Posting to the ERP or accounting system, or deleting records and files.
- Exceptions the Skill was told to flag.
In practice: set write tools on connectors to Needs approval or Blocked, and use Manual mode in Cowork for sensitive files or accounts. Anthropic’s safety guidance also says not to schedule tasks that send messages on your behalf or make purchases.
Security and data
- Business or consumer plan. On the consumer plans (Free, Pro, Max, Ultra) you choose whether your chats can be used to train Anthropic’s models. Team and Enterprise fall under the Commercial Terms: by default, inputs and outputs are not used for training, except feedback a user deliberately sends. For company work, use a business plan.
- GDPR. Under the Commercial Terms, your company is the controller and Anthropic the processor. Check the data processing terms and where data is stored with your data protection lead before connecting systems that hold personal data.
- What not to hand over. Anthropic advises against giving Cowork access to local files with sensitive information, such as financial documents, and suggests a dedicated working folder instead of broad access. Keep payroll, bank details and personal data out unless the task needs them.
- Skills from elsewhere. A Skill can contain scripts and instructions, so a malicious one can attempt prompt injection or send data out. Install Skills only from sources you trust and read any Skill before you turn it on, including ones from colleagues.
- Admin controls. On Team and Enterprise plans, owners can turn Cowork off, stop members using automatic approval and set connector permissions for everyone; Enterprise can also scan Skills for malicious content and retrieve Cowork sessions through the Compliance API.
Your AI policy should say which of these settings you use and who can install Skills and connectors. This is not legal advice.
Where this stops
Claude working through a Skill on your exports and shared drive is still an assistant: it reads what it is given and produces a draft for a person. That is not the same as software that writes back to your ERP with approvals, an audit trail of who approved what, and retries when a system is down. Connectors can take actions, but only where the connector offers them and only as far as the permissions you set.
A useful line: if a Skill runs a few times a week and a person checks every result, keep it as a Skill. If it runs on every invoice and its output should land in the ERP, it has become a workflow and should be built as one, with the ERP as the source of truth. Our guide to AI for ERP covers the options.
If you are deciding which procedure to turn into a Skill first, the free Bonofici grader reads your company’s public pages and shows where to start.