Claude, ChatGPT, Copilot and Gemini: What a Team Learns Differently

Training a team on Claude, ChatGPT, Copilot or Gemini is mostly one shared skill and one product-specific one. The shared skill is prompting, and it now transfers almost completely between the four. The product-specific part is the working model each assistant asks a person to carry in their head: what the reusable unit is, where context comes from, and which failure the tool makes easy. Get the second part wrong and people use an expensive assistant as a faster search box.

Published 5 September 2026. Last updated 5 September 2026. Every capability claim below links to the vendor page it came from, each one read on that date.

What has to be taught differently for each assistant?

Four things, and only four. The unit of reuse: the thing a person builds once and returns to. The source of context: whether the assistant is fed material deliberately or reaches into the organisation on its own. The strongest real use: where the tool earns its licence rather than where the marketing points. And the first mistake: the specific misunderstanding that a new user arrives with, which is different for each of the four and, in three of the four, is not a prompting problem at all.

This page is deliberately not a feature comparison, and it is not a data-privacy page. Training defaults, EU residency, retention and admin controls all sit in a separate reference with every cell sourced to the vendor doc: does your AI vendor train on your data? Read that one before you decide which account type people sign in with. Read this one after, when you have to make the tool actually useful.

The unit of reuse, side by side

Claude: a container of context, and a written procedure

Anthropic gives a team two reusable objects, and they are not the same shape. A project is a container: Anthropic describes projects as workspaces “with their own chat histories and knowledge bases,” where “you can upload documents, text, code, or other files to a project's knowledge base” and “define project instructions for each project to further tailor Claude's responses.” The natural mapping is to a body of context: a client, a case file, a campaign, a product line. It is not a persona.

A skill is a procedure. Anthropic describes Skills as “reusable, filesystem-based resources that give Claude domain-specific expertise: workflows, context, and best practices that turn a general-purpose agent into a specialist,” and draws the distinction that matters for training: “unlike prompts (conversation-level instructions for one-off tasks), Skills load on demand, so you don't have to repeat the same guidance across conversations.” Each one is a SKILL.md file with a name and a description, and the documentation is specific about what the description has to do: it “must include both what the Skill does and when Claude should use it.”

That is a genuinely teachable act for a non-engineer, because it is just writing down how your organisation does a thing and when it applies. The strongest practical use is long-document work with a stable body of reference material behind it: a contract set, a policy library, a research corpus.

ChatGPT: the configured prompt is the asset

OpenAI's own documentation describes the reusable unit as a shape rather than a container. A well-formed instruction set has four parts in sequence: identity (purpose, communication style and goals), instructions (the rules), examples (sample inputs paired with desired outputs) and context (the supporting material). There is a priority order underneath it that people rarely know: developer messages sit above user messages, so the standing rules outrank whatever someone types next. And the formatting advice is refreshingly plain: “markdown headers and lists can be helpful to mark distinct sections of a prompt, and to communicate hierarchy to the model.”

Grounding an answer in your own material has a name in the same guide. OpenAI writes that “the technique of adding additional relevant context to the model generation request is sometimes called retrieval-augmented generation (RAG),” and points at file search as the built-in route. For a business team the teachable version is: decide what the assistant is allowed to answer from, then supply it, rather than hoping the model remembers.

The most useful 2026 change is a governance one. OpenAI is now steering teams away from storing prompts inside the platform: “OpenAI is deprecating reusable prompt objects in the API. Prompt creation will be de-emphasized beginning June 3, 2026, and v1/prompts is scheduled to shut down on November 30, 2026.” The recommendation is to “store production prompts in your application code instead,” because “code-managed prompts let you use typed inputs, code review, tests, and your normal deployment process to change model behavior.” A finance or legal team has no application code, but it has the same problem and the same answer: the prompt that matters belongs in a shared document with an owner and a change history.

Gemini: value comes from already being where the document is

Gemini's reusable unit is a Gem, and Google's own instructions for writing one are a compact prompting curriculum: cover persona, task, context and format, and “provide details about your goals, desired behaviors, and preferred format to help the Gem personalize Gemini's responses.” Files go under a Knowledge section, and Google is direct about why: “you can add files to give your Gem more context or to reference specific docs in your chats with the Gem.”

The real advantage is positional rather than architectural. Google describes Gemini in Workspace as reaching the material in place: it “accesses customer data in order to provide personalized responses, such as summarizing a document in Google Docs or analyzing data in a Google Sheet.” The permission rule is the same one Microsoft applies, and stated just as flatly: “if the user doesn't have access to a document or email, Gemini will not retrieve that content.” For an organisation already living in Docs, Sheets and Gmail, the training is mostly about habit, not technique. People do not need to be taught to paste; they need to be taught to stop pasting.

One documented exception belongs in every Workspace induction, because it is the surface most likely to be handed a folder of internal material. Google records that when files are pulled into Gemini Notebook, “your organization's file sharing and data region settings do not apply to data in Gemini Notebook.” That is not a reason to ban it. It is a reason to name it in the policy, alongside the admin controls Google documents for retention and for turning the surface on or off.

Why is permissions hygiene the real Copilot training?

Because Copilot is the only one of the four whose context comes from the organisation by default rather than from the person. Microsoft describes the product as coordinating large language models with “content in Microsoft Graph, such as emails, chats, and documents that you have permission to access,” and summarises the grounding rule as “access scoped by user permissions (security and compliance enforced).” The privacy documentation makes the consequence explicit:

“Microsoft Copilot only surfaces organizational data to which individual users have at least view permissions. It's important that you're using the permission models available in Microsoft 365 services, such as SharePoint, to help ensure the right users or groups have the right access to the right content within your organization.”

Read the second sentence as an instruction rather than a disclaimer. Microsoft adds that this “includes permissions you give to users outside your organization through inter-tenant collaboration solutions, such as shared channels in Microsoft Teams,” and elsewhere that the semantic index “honors the user identity-based access boundary so that the grounding process only accesses content that the current user is authorized to access.” Nothing leaks that was not already reachable. What changes is the effort. A document buried three levels down in a site nobody visits was protected by obscurity, and an assistant that searches the whole tenant in a second removes exactly that protection.

Microsoft's own readiness guide (Microsoft Learn, get ready for Copilot) is unusually candid about where the exposure comes from, and it contains the single most useful sentence for anyone planning a rollout: “by default, SharePoint sets sharing settings to the most permissive option.” The same guide frames the preparation as content governance rather than AI work, with steps to “identify potentially overshared content,” “define Copilot readiness for the organization,” and run a report that identifies “the top 100 sites where content was shared with your entire organization in the past 28 days.” There is a control for the sites you cannot fix in time: restricted content discovery “prevent[s] content from appearing in Copilot or agentic experiences and in organization-wide search queries” and “reduce[s] accidental exposure while leaving site permissions unchanged.”

So the honest curriculum for a Copilot organisation is roughly two hours of prompting and a permissions project that starts before the licences do. Curricula built around prompting alone leave that half untouched, which is how a rollout ends up generating complaints about what the assistant found rather than about what it wrote. If your organisation is also thinking about what an AI literacy programme has to contain under Article 4 of the EU AI Act, this is the piece of it that produces evidence a regulator would recognise: a documented access review, dated, with actions attached.

What actually transfers in prompting in 2026?

Five moves, and both major vendors document versions of all of them. This is the part to teach once, to everyone, regardless of which assistant they have been issued.

1. Assume no shared context. Anthropic's framing is the best available for a non-technical audience: “think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.” It comes with a test anyone can apply: “show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too.”

2. Give the reason, not only the rule. Anthropic's worked example is one line long. Less effective: NEVER use ellipses. More effective: Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them. The documentation explains why this works at all: the model “is smart enough to generalize from the explanation.” Teach it as: state what the output is for.

3. Show, do not adjectivise. Anthropic calls examples “one of the most reliable ways to steer Claude's output format, tone, and structure” and recommends including “3–5 examples for best results,” relevant, diverse and wrapped in tags. OpenAI recommends showing “a diverse range of possible inputs with the desired outputs.” Two good examples beat a paragraph of adjectives, every time.

4. Separate the parts. Anthropic: “XML tags help Claude parse complex prompts unambiguously, especially when your prompt mixes instructions, context, examples, and variable inputs.” OpenAI reaches the same place through markdown headings. The syntax is not the lesson. The lesson is that instructions, context, the material and the requested format should be four visibly separate blocks, whichever notation the person prefers.

5. Put the long document first, and make it quote. For inputs around 20k tokens and above, Anthropic's guidance is to “place your long documents and inputs near the top of your prompt, above your query, instructions, and examples,” noting that “queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs.” Pair it with the grounding move: “for long document tasks, ask Claude to quote relevant parts of the documents first before carrying out its task.” That single habit is the practical answer to “I pasted a forty-page contract and it invented a clause,” because a fabricated quote is checkable with a find command and a fabricated conclusion is not.

Where the tools genuinely diverge is model choice, and that is now part of the skill. OpenAI splits its own range in two: it trained the o-series models (“the planners”) “to think longer and harder about complex tasks, making them effective at strategizing, planning solutions to complex problems, and making decisions based on large volumes of ambiguous information,” while “our lower-latency, more cost-efficient GPT models (‘the workhorses’) are designed for straightforward execution.” Its selection rule is plain: if speed and cost matter most and the tasks are “straightforward, well defined,” use the second kind. The teachable version: pick the model class first, then match prompt density to it. High-level goal for a reasoning model, explicit specification for a fast one.

Which prompting advice has expired?

A good deal of prompt training still teaches the 2023 repertoire. Several parts of it are now either useless or actively harmful, and in each case the vendor documentation says so.

The last row is the one that changes behaviour in real organisations, because it inverts a habit people were explicitly taught. Anthropic's migration guidance for current models is to remove the pressure language: “where you might have said ‘CRITICAL: You MUST use this tool when…’, you can use more normal prompting like ‘Use this tool when…’.” Prompts written for older models overtrigger on newer ones. A team that inherited a prompt library in 2024 is now carrying instructions that make results worse, and nobody has told them.

The “take a deep breath” family deserves a specific note, because it is still repeated in courses. It is not that the technique was disproven; it is that it appears nowhere in either vendor's current prompt-engineering documentation, which is the only source a training programme should be teaching from. Advice with no current primary source behind it is folklore, and folklore is expensive to un-teach.

What does a new user get wrong first on each?

Notice that three of the four are not prompting mistakes at all. They are mistakes about where things live: context, configuration, policy and permission. That is the structural reason a generic prompt-engineering course underperforms on a real rollout, and it is the same reason a plain-language explanation of what these systems are tends to do more good in week one than a list of prompt tricks.

How to sequence a rollout across two tools

Ending up with two assistants rather than one is a common shape: a productivity suite ships with one, and a team buys the other for a specific job. That is workable, and it is cheaper to teach than it looks, provided the sequence is right.

First, settle the account type. Which identity people sign in with decides the vendor's training default, the retention posture and the admin surface on every product in this comparison. That is a procurement decision with a documented answer per tier, set out in the tier-by-tier vendor reference, and it should be closed before anyone books a session.

Second, do the permissions work if Copilot is in scope. Microsoft's readiness guide is the checklist, and it produces a dated artefact you can put in an evidence file.

Third, teach prompting once, for everyone. The five moves above, with the expired advice named explicitly so people stop repeating it.

Fourth, teach the unit of reuse per tool, briefly. Thirty minutes each. What a project is, what a skill is, what a Gem is, what a configured prompt is, and where each of them is allowed to live.

Fifth, write the content rule. Which classes of material may go into which surface. This is the part that survives a staff change, and it is also the part an Article 4 literacy programme can point at as a measure appropriate to the systems you deploy. If you also run AI systems that speak to customers, the transparency duties are a separate obligation with separate deadlines: see the EU AI Act compliance checklist and whether an AI caller must identify itself.

For the everyday layer underneath all of this, our guide to practical business uses of ChatGPT covers what teams reach for first, and the AI glossary is a reasonable handout for people who nod along to words they have not been given. If your organisation works in a language that is not English, the quality picture changes in ways none of these four vendors publish figures for: see AI in smaller European languages.

What this page does not claim

It does not rank the four. A ranking would depend on which suite the organisation already runs, which is not a fact about the assistants. It does not repeat the data-handling tables from the vendor reference, and it does not restate the EU residency comparison or the vendor security assessment questions, all of which are maintained separately and re-verified on their own schedule.

It also does not quote anything that could not be read on a vendor page. The one gap is named above: OpenAI's consumer help centre blocks automated retrieval, so this page draws its ChatGPT material from OpenAI's developer documentation instead. Vendor documentation moves quickly, sometimes without a changelog, and every page cited here carries the date it was read. If a quotation no longer matches the page it links to, the link is there so you can find that out before we do.