Three pieces make AI useful inside a business.
A Project is where AI learns how one part of your business works. A Skill is a repeatable job it knows how to perform. A Connection (Claude) / Plugin (ChatGPT) is how it reads and acts in the systems you already run.
Without the first two, every chat starts from zero. The person asking has to re-explain the site list, the supplier terms, what "available" means in your stock system, and the rule about never raising a purchase order without approval. Without the third, the AI can reason about the work but a person still exports the CSV and keys the result back in.
This is a reference for setting up all three.
Projects
A Project is a dedicated workspace for one part of your business. It holds the instructions, files, definitions and ongoing conversations for that team, so context is established once rather than in every chat.
What changes
Asked in a plain chat, "find stock we should move between stores" returns a generic answer. The AI knows what you asked but nothing about how you decide.
Asked inside a Buying and Inventory Project, the same question is interpreted against rules the workspace already holds:
- Prefer warehouse stock first.
- Check what transfers are already underway.
- Account for incoming purchase orders.
- Never leave the supplying site below its minimum buffer.
- Always show proposed transfers before creating them.
What you get
- Consistency. Everyone in the department works from the same rules and definitions, so two people asking the same question get the same answer.
- Less repetition. Context is written once, not retyped per chat.
- Safer actions. Rules such as never create a purchase order without approval live in the workspace, not in whoever remembers them.
- Organisation. Finance work stays with Finance, buying work with Buying.
- Continuity. Related decisions stay together instead of scattering across one-off conversations.
How to structure them
One Project per department is the usual shape.
Leadership. Used by owners and senior management. Holds business goals, KPI definitions, the site list, reporting preferences, key suppliers and financial targets.
Buying and Inventory. Used by buyers, inventory and operations. Holds the supplier list, brand and category structure, SKU conventions, stock and transfer rules, buying calendars and lead times.
Finance. Used by finance and accounts. Holds chart-of-accounts guidance, reporting definitions, invoice and payment processes, credit-note rules and approval limits.
Operations. Used by site and area managers. Holds the site list, transfer and adjustment procedures, daily routines, operational KPIs and escalation rules.
The test for whether something belongs in a Project: if it has been explained in a chat more than twice, it should be written down there.
Skills
A Project holds context. A Skill holds procedure.
A Skill captures a specific piece of recurring work. A complete one specifies six things:
- When to use it. The trigger or question that should invoke it.
- What it needs. The information required before it can run.
- Which systems it reads. The connections it draws on, named explicitly.
- The steps. The sequence to follow, in order.
- The business rules. Your constraints, stated as rules rather than preferences.
- The output and the approvals. What the result should look like, and what requires a person to say yes before it happens.
A worked example
A stock risk Skill is not a prompt. It is a written procedure:
Use when: identifying products at risk of running out of stock.
Steps: Check current stock levels. Review recent sales activity. Check stock across every site. Check incoming purchase orders. Identify high-risk products based on sell-through against stock on hand. Recommend transfer, wait for incoming stock, reorder, or monitor.
Output: Group findings under Urgent, Transfer, Incoming, Reorder and Watch, with reasoning for each.
Approval: Always ask before acting. Where transfers are recommended, create them only after explicit approval.
Written this way, the job runs identically in March and in December, and whether the person running it has been there ten years or ten days.
Common Skill categories
Most businesses find their first Skills fall into five groups.
- Risk. What are we about to run out of, and what should we do about it?
- Rebalancing. What is in the wrong location, and where should it go?
- Buying. What needs ordering, and what should the draft order contain?
- Trading. What is selling now, what is accelerating, what has gone quiet?
- Reporting. The recurring summary that draws on two or three systems and lands as five headings rather than fifty rows.
Connections and Plugins
Projects supply context and Skills supply procedure. Neither produces work without access to the systems where the work lives. Claude calls that access a Connection and ChatGPT calls it a Plugin. The two work the same way.
What a connection covers
Inventory and ERP. Read stock by location, review sales activity, check incoming purchase orders, review existing transfers, retrieve product and supplier records. On approval, create transfers, purchase orders and adjustments.
Accounting. Retrieve reports, list outstanding invoices, review bank transactions and payments, compare credit notes across systems and flag inconsistencies. On approval, create invoices, payments and bank transactions.
CRM, POS, ecommerce and marketing. The same principle applies wherever the work lives.
The read and write rule
Two rules keep this safe.
Reading is unrestricted. Any question, any time, across every connected system.
Writing waits for a person. Every transfer, purchase order, invoice, payment and credit note is proposed as a draft and created only after someone explicitly approves it.
That distinction is what allows teams to move quickly. An assistant that reads widely and asks before it acts does not require supervision.
Identifying your next Skill
The best Skills already exist as work someone does manually. Look for tasks performed:
- Every day, every week, or every month.
- Before every meeting.
- Whenever a particular issue occurs.
- By copying information between systems.
- By opening multiple reports.
- By checking one system and then another.
- By checking a system and then sending an email.
The question that surfaces them fastest: what do I repeatedly do that requires me to gather information, apply the same reasoning, and then take an action?
Gather, reason, act. That shape is what a Skill is for.
Turning a workflow into a Skill
Skills are captured after the fact, not written up front.
1. Do the real task. Work through an actual job step by step, in the real systems. For example: review every overdue purchase order and determine which suppliers need chasing.
2. State what you would change. This is where the business rules surface. Exclude anything less than three days overdue. Group by supplier. Put urgent issues at the top. Give the reasoning, not only the instruction.
3. Capture it. We do this every week. Turn the process we just followed into a reusable Skill. It should capture all six elements listed above.
4. Run it and keep editing. Skills are not finished artefacts. They improve each time someone notices the output missed something.
What this produces
Context written once. Procedures that run the same way every time. Real access to the systems where work happens, with a person on every action that changes something.
The result is that AI stops being a place people go to ask questions and becomes part of how work gets done. The recurring summary assembles itself and it is correct. Stock risk is flagged on Tuesday rather than discovered on Friday. The purchase order arrives as a draft, with its reasoning attached, waiting on a yes.
The effect is not on headcount. It is that the people paid for their judgement stop losing half their week to assembling context, and get it back for the decisions that need a person in the room.