How to Create Skills in ChatGPT and Claude

A lot of people still use AI in a loop: write a prompt, get a result, rewrite the prompt, and repeat the same explanation next time. Some keep their best prompts in a notes app and paste them into every new conversation.

That works, but it does not scale particularly well. A skill gives you a better option.

In ChatGPT and Claude, a skill is a reusable workflow for a particular job. It can contain detailed instructions, examples, reference files, documentation, templates and a defined output format. You decide which parts must remain fixed and where the AI is allowed to use judgement.

The result is not identical content every time. The result is a consistent standard every time.

01What an AI skill actually does

Think of a skill as a playbook that the AI can load when a relevant task appears. Instead of explaining your process from scratch, you define it once and let the assistant apply it to new inputs.

A good skill can specify:

• when the workflow should be used;
• what information it expects to receive;
• the steps, rules and checks it should follow;
• reference material or examples it may need;
• the exact format of the finished output.

That makes skills particularly useful for recurring work: inbox triage, investment memo reviews, meeting preparation, weekly reporting, document creation, research summaries and quality-control checks.

02The four-question framework

Before creating a skill, answer four questions:

1. What request should trigger it? Describe the task or situation clearly enough for the AI to recognise when the skill is relevant.
2. What information will it receive? Define the emails, documents, data, user instructions or connected sources that provide the input.
3. What should it do with that information? Set out the process, decision rules, priorities and quality checks.
4. What should the output look like? Specify the headings, fields, order, tone, length and format you want.

These questions stop a skill from becoming a vague collection of prompt fragments. They turn it into a repeatable workflow with a clear start and finish.

03Example: an email triage skill

Suppose you want ChatGPT to summarise the last 24 hours of email in a way that is genuinely actionable. First, connect the relevant inbox through your ChatGPT settings, subject to the access available on your plan and workspace. The connection provides the emails; the skill provides the rules for what to do with them.

You might ask ChatGPT to create a skill with the following logic:

• High priority — ASAP: messages that explicitly require an immediate decision or action from you.
• High priority: messages sent directly to you by someone in your organisation, or active conversations in which you are already involved.
• Medium priority: general internal messages, conversations where you are copied, and relevant third-party outreach.
• Low priority: newsletters, generic marketing and other non-actionable messages.

If a message matches more than one category, the highest applicable priority wins. That one rule removes ambiguity and keeps the summary useful.

04Define the output, not just the task

The most important step is often the one people skip: describing the finished result.

For the email example, ask for a consistent table with these columns:

Tell the skill not to invent an action or deadline when one is missing. Ask for concise, professional wording and place the most important messages first.

Now the content will change with every inbox, but the categories, standards and structure will remain consistent.

05Build, test and refine

Once the skill is installed, test it on real inputs. In this example, the test request is simply: “Summarise my emails.” ChatGPT can recognise that the task matches the installed skill and apply the triage workflow. Claude follows the same general pattern, loading relevant skills when a request matches their description.

Do not treat the first version as final. If the result is too rigid, allow more judgement in a specific part of the process. If it is too flexible, add a rule, an example or a stricter output format. If it triggers at the wrong time, improve its name and description.

The test is straightforward: run it on different sets of real input and check whether the standard holds.

06Skills versus saved prompts

A saved prompt is useful for a one-off instruction you want to reuse manually. A skill is better when the job involves a procedure, supporting material, multiple steps or a fixed quality bar.

Skills also make good work easier to share. Both OpenAI and Anthropic support the open Agent Skills format, so a well-designed workflow can be moved between compatible products. The exact installation, connected accounts and tool permissions may still need to be configured separately.

Decide what must remain consistent and where AI can use its judgement. Build the workflow once, test it on real inputs, then reuse it.

07Start with work you already repeat

The best first skill is usually not an ambitious new AI project. It is a task you already perform regularly and can explain clearly.

Look for a prompt you keep rewriting, a checklist you follow from memory, or a deliverable whose structure barely changes. Then answer the four questions: trigger, input, process and output.

That is enough to turn a useful prompt into a reusable workflow — and to make ChatGPT or Claude behave less like a blank chat box and more like an assistant that understands how you work.