AI Skills for business: from one-off prompts to repeatable work

Most businesses do not have a shortage of prompts. They have a shortage of dependable ways to repeat good work.

Someone finds a useful approach in ChatGPT, saves it in a document, shares it with a colleague and assumes the problem is solved. A few weeks later, three versions of the prompt exist, nobody remembers which source material is current and the checking step has quietly disappeared.

AI Skills are interesting because they can package more than a clever instruction. They can hold a defined workflow, examples, supporting resources and, where the platform permits it, code. The useful idea is not that a Skill makes the underlying model more intelligent. It makes a particular way of working easier to repeat.

What is an AI Skill?

OpenAI describes Skills as reusable, shareable workflows that help ChatGPT complete specific tasks more consistently. A Skill can contain instructions, examples, supporting resources, repeatable steps and code. Once installed, ChatGPT can use one or more relevant Skills when they are helpful.

That makes a Skill closer to a small operating manual than a saved prompt.

A useful business Skill might define:

when the workflow should be used;
which information is required;
which sources are authoritative;
the steps to follow;
the required output structure;
what must be checked by a person;
which claims or fields must never be changed;
what to do when information is missing or uncertain.

Skills, prompts, apps, plugins and agents are not the same thing

These terms are often collapsed into one vague description of an AI tool. They serve different purposes.

Capability	What it is useful for	What it does not provide by itself
Prompt	Giving instructions for one task or conversation	A maintained operating process
Skill	Packaging reusable instructions, examples and workflow guidance	Permission to access an external service
App	Connecting an AI product to external information or supported actions	The complete method for doing the work well
Product plugin	Packaging capabilities using one provider's extension system	Portability to every other provider
Agent Plugin	Packaging portable Skills and MCP server definitions in a standard directory	Shared permissions, sandboxing or trust across clients
Agent	Pursuing a goal using instructions, tools, files and other capabilities	Guaranteed accuracy or unrestricted autonomy

OpenAI's current product guidance makes one separation particularly clear. Skills provide reusable instructions and workflow guidance. Apps connect external accounts, information and actions. A ChatGPT plugin can package either or both, but installing it does not bypass the permissions of the connected service.

There is now a separate open standard called Agent Plugins 1.0. It defines a portable directory for Agent Skills and MCP server definitions. The name is similar, but the scope is different. The complete Agent Plugin explanation covers what the format makes portable and what remains specific to each client.

The practical business value

The business case is not “we can install more AI”. It is that useful knowledge can move out of one person's chat history and into a reviewed, reusable process.

Preserve a good method

A proposal-review Skill could require the user to identify the decision, read approved source documents, separate facts from assumptions, flag missing evidence and return a consistent review table. That is more valuable than a prompt saying “review this proposal”.

Protect brand and factual boundaries

A content Skill can define tone, structure and preferred explanations while protecting names, dates, prices, citations and legal wording from stylistic rewriting. The Skill does not guarantee compliance, but it can make the checking boundary explicit every time.

Make training reusable

Training normally fades if it remains a presentation. A Skill can turn the method taught in a workshop into something staff can invoke during real work, with the same steps, examples and review rules available afterwards.

Combine specialist capabilities

More than one Skill can contribute to a task. A business might combine its own report structure, a brand voice method and a spreadsheet-analysis capability. The important word is combine. One enormous Skill that tries to govern every task becomes difficult to understand, test and maintain.

A sensible first Skill for a Scottish SME

Start with a task that already has a recognisable method and a person who can judge the result.

Good candidates include:

turning meeting notes into actions using an agreed format;
checking a document against a standard checklist;
preparing a customer-enquiry draft from approved service information;
converting one verified source asset into channel-specific content drafts;
producing a weekly report from a defined set of figures and source notes;
assessing whether an automation idea is suitable for AI, ordinary rules or neither.

Avoid starting with legal decisions, staff assessment, financial commitments, unsupervised publishing or a workflow that can change customer records without approval.

Do not brainstorm Skills. Look for repeated work

The wrong way to choose a business Skill is to sit in a meeting and make a list of everything AI might do.

Start with evidence from the work already happening. Look for tasks people repeat, methods that produce a good result, places where staff regularly get stuck and useful approaches that live inside one person's chat history.

Anthropic's Smart Reports beta, announced on 10 September 2026 for Claude Enterprise, is an interesting example of this shift. Anthropic says the reports analyse how teams use Claude, including what work is being completed, what it costs, where sessions encounter friction and which repeated patterns may be worth packaging as shared Skills.

The useful idea is broader than one Claude feature. A business can carry out the same review manually:

collect a small sample of recurring AI-assisted tasks;
group them by outcome rather than department or tool;
identify the tasks with a recognisable successful method;
record where users repeat instructions or corrections;
separate missing training from a missing workflow;
choose one pattern that is valuable, frequent and safe enough to test;
package the method, then compare it with the original way of working.

Frequency alone is not enough. A task repeated every day may still be a poor Skill if the inputs are uncontrolled, the judgement cannot be explained or the consequence of an error is too high.

The Data plugin shows how Skills fit around business systems

OpenAI released its Data plugin for ChatGPT Work and Codex on 10 September 2026. It can analyse connected business data, investigate changes and create reports or dashboards. OpenAI's guidance recommends connecting authoritative metric definitions through a semantic layer and checking sources, periods, filters and definitions before relying on a result.

The important point for Skills is that the plugin can be customised with templates and context Skills. A team could package its approved definitions, analysis method, reporting structure or dashboard design guidance so the Data plugin applies them to future work.

That is a useful division of responsibility:

the data source provides the records the user is permitted to access;
the semantic layer provides agreed business definitions;
the plugin provides the analysis capability;
the Skill provides the organisation's method and context;
deterministic checks validate fixed calculations or required fields;
a person reviews the conclusion and any consequential action.

This is what practical AI adoption looks like. The Skill is not a replacement for the data warehouse, permissions, business definitions or human judgement. It is the reusable layer that tells the wider system how this organisation expects the work to be done.

For a smaller Scottish business without a data warehouse, the same architecture can begin with an approved spreadsheet or exported report. The technology can be simpler while the principles remain the same: known sources, clear definitions, a repeatable method and visible checking.

How to design one properly

1. Write down the current method

Observe how the work is actually completed. Capture the inputs, judgement points, exceptions and checking steps. Do not automate the neat version described in a meeting if the real process behaves differently.

2. Define the boundary

State what the Skill does, what it must not do and when it should stop. A narrow Skill is easier to test than a general instruction to “manage our marketing” or “run customer service”.

3. Separate facts from expression

Identify protected facts, approved sources and fixed fields before adding tone or style guidance. A natural voice is useful. A naturally written false claim is still false.

4. Test awkward examples

Use missing information, contradictory instructions, outdated source material, unusual formatting and cases that require escalation. Record the expected behaviour before judging the output.

5. Assign an owner

Someone must review source changes, update examples, decide when the workflow no longer fits and retire old versions. Reusable does not mean maintenance-free.

Security and trust matter

OpenAI advises users to review an uploaded Skill and trust its source because a Skill can include instructions, supporting files and code. OpenAI scans uploaded Skills, but its guidance says that scanning should not replace the organisation's own review, policies or judgement.

For a business, review should cover:

who created and maintains the Skill;
which files, instructions and scripts it contains;
which apps or external services it expects to use;
which read and write actions those connections permit;
which information staff may provide;
where human approval is required;
how updates and versions are controlled.

A Skill should never be treated as a way round normal data protection, security, procurement or professional obligations.

Are Skills portable between ChatGPT, Claude and Gemini?

The broader pattern is becoming visible across providers, but the implementations are not identical.

OpenAI supports reusable Skills in eligible ChatGPT workspaces and in other supported OpenAI products, with availability varying by plan and product surface. Anthropic documents custom Agent Skills for Claude as bundles built around a SKILL.md file, with relevant Skills loaded when required and multiple Skills able to work together.

Google introduced user-facing Gemini Skills on 30 September 2026. They can be created from chats, invoked with a slash command, selected automatically, combined and supplied with reference files. Google says these Skills will replace Gems, with Workspace business, enterprise and nonprofit Gems due to end from March 2027. Google also documents modular Skills in its developer tooling, including ADK examples. The shared name does not make the user-facing and developer implementations identical.

The Agent Plugins 1.0 standard is a meaningful step towards portability. It gives compatible clients one predictable package structure for Agent Skills and MCP server definitions. It does not make the clients identical.

Distribution, installation, permissions, sandboxing, publisher trust and client-specific capabilities remain outside the portable core. A package may load in more than one supported client while its available connections, approval flows and execution behaviour still differ.

Businesses should preserve the provider-neutral method, source material and tests, use the shared package format where appropriate, and review the runtime controls in every client where it will be installed.

The Scottish opportunity

Scotland's AI strategy for 2026 to 2031 emphasises responsible and inclusive adoption that produces tangible benefits for businesses and communities. The National AI Adoption Programme also focuses on moving Scottish SMEs from awareness towards practical implementation.

Skills fit that gap well when they capture a real method. They give a small organisation a way to reuse specialist knowledge without pretending it has built a fully autonomous system. They can help turn training into practice, practice into a workflow and a proven workflow into a repeatable capability.

My view is simple: the valuable product is not the prompt. It is the judgement, structure, safeguards and repeatable outcome packaged around it.

Sources and further reading

See how a business can use ChatGPT Skills safely →