ChatGPT Skills: What They Are, How to Use Them, and How to Build Your Own

So you have finally written the perfect prompt. It understands the job, follows your format, remembers the annoying little rules and produces something genuinely useful.

Then you open a new conversation and have to paste the whole thing again....and again.

(unless you have things written in memory!)

That is the problem Skills are meant to solve.

A Skill is a reusable package of instructions and supporting files that teaches ChatGPT or Codex how to handle a particular kind of task. It can contain a workflow, examples, reference material, templates, images and, when needed, scripts.

Instead of explaining your process from scratch every time, you build it once and invoke it again with @Skill Name.

Skills are useful, but they are not magic. They do not train a private model, unlock unavailable tools, bypass policies or guarantee perfect results. We will get to the limitations before building anything, because discovering them after two hours of setup is how laptops learn to fly.

Current as of 7 August 2026: Skills are developing quickly. Interface names and availability may change. Check the official OpenAI Skills guide if your app looks different from the screenshots in this guide.

❓ What is a Skill?

A Skill is a folder built around one required file called SKILL.md. That file contains the Skill's name, a description that helps ChatGPT recognise when to use it, and the instructions for completing the task.

The folder can also contain:

References: detailed guidance, policies, examples, schemas or background information.
Assets: images, logos, fonts, templates and other files used while producing the result.
Scripts: small programs for tasks that need reliable, repeatable execution.
Interface details: the display name, description and default prompt shown in ChatGPT.

You do not need to understand the folder structure to create a Skill. @skill-creator can build and validate it for you. Understanding the structure helps when you want to make a Skill more reliable or share it with other people.

🧠 Skills, prompts, memory, Projects and plugins are not the same thing

These features overlap, which is why discussions about them often become a bowl of terminology soup.

A writing Skill might teach ChatGPT how to produce articles in your preferred structure. Memory might store that you dislike corporate language. A Project might contain the research and previous articles. A plugin might connect the workflow to Google Drive or another service.

They can work together. They are not interchangeable.

🤔 What can you use Skills for?

Good Skills handle jobs where the method matters repeatedly.

Examples include:

writing articles in a defined voice and structure;
turning meeting notes into actions, owners and deadlines;
checking documents against a brand or accessibility standard;
producing a weekly project update from the same kinds of sources;
reviewing pull requests using a team's conventions;
creating presentation decks from a house template;
organising research into a consistent evidence format;
generating recurring characters using curated visual references;
preparing social posts for several platforms without rewriting the rules each time.

“Answer every question brilliantly” is not a useful Skill. “Turn messy project notes into a one-page stakeholder update with decisions, risks, owners and next actions” is.

A Skill should have a recognisable job and a result you can judge.

Where Skills work

OpenAI currently documents standalone Skills for the ChatGPT desktop app, Codex CLI and the Codex IDE extension. Skills packaged inside plugins can also run on supported plugin surfaces, including ChatGPT Work on the web. The current surface list is maintained in the official Build Skills guide.

In ChatGPT, Skills are selected with @. Codex uses its own Skill picker or $ mentions depending on the surface.

Web

On the web, Skills and plugins are centred on ChatGPT Work, not an ordinary lightweight Chat conversation. If a Skill does not appear in your usual chat, switch to Work and check the Skills section under Plugins.

Desktop app

The ChatGPT desktop app supports standalone Skills. Open the Skills area in the sidebar to view installed Skills, then use @ in a supported conversation to select one.

Phone app

The native mobile Chat experience does not currently offer the same standalone Skill workflow as desktop.

You can continue a Cloud Work conversation from mobile. Remote can also connect to a desktop host, allowing the work to run where the Skills and configuration are available while you steer it from your phone. That is not the same as opening a fresh normal mobile chat and invoking any standalone Skill directly.

Codex

Codex can use Skills in its desktop, CLI and IDE workflows. Local and repository Skills may have different scopes, so a Skill available inside one project may not appear everywhere.

Availability can also depend on your plan, account, workspace settings, administrator controls and rollout region.

✖️ The limitations, without the marketing perfume (you can skip it but reccomending reading before creating skills!)

Skills make workflows reusable. They do not make ChatGPT omnipotent.

A Skill cannot create a capability that the host does not have

A Skill can teach ChatGPT how to use image generation, a browser or a connected service. It cannot make those tools appear on an account or surface where they are unavailable.

Instructions saying “send this to Slack” do not create a Slack connection. A connector or plugin tool must exist and the account must have permission to use it.

A Skill does not bypass permissions, approvals or safety policies

The Skill operates inside the same permission and policy boundaries as the conversation. It cannot quietly grant itself access to private files, external accounts or restricted actions.

Scripts may still require an environment where they can run. Network access, file access and external actions may be blocked or require approval.

A Skill is not model training

Bundling reference photographs, writing samples or examples does not fine-tune the model. The Skill supplies those files as reusable working context.

That can improve consistency substantially, but it does not produce a permanent LoRA-style identity model or guarantee an identical face in every generated image.

Activation is not guaranteed

ChatGPT can invoke a Skill automatically when the request matches its description, but automatic matching may fail. Descriptions can also be shortened when many Skills are installed.

Use an explicit @Skill Name mention when the Skill matters.

Instructions can still conflict

The current request, the Skill, other active Skills, available tools and platform policies all influence the result. Vague or contradictory instructions will not become clear merely because they live inside a folder.

More instructions can make a Skill worse

A 9,000-word Skill stuffed with every thought its creator has ever had can bury the workflow it was supposed to protect.

Keep the main instructions short. Put detailed material in separate reference files and tell ChatGPT when to read each one.

Results remain probabilistic

A good Skill improves process consistency. It does not remove model variation, image artefacts, misunderstood requests or ordinary mistakes.

Build a checking step into the workflow. A Skill that generates something and never inspects it is an assembly line with no quality control.

Shared Skills can expose bundled material

Do not place passwords, API keys, private documents, confidential client information or intimate personal photographs inside a Skill you plan to share.

Create a clean public template with placeholders. Let each person build a separate private copy containing their own material.

Platform support is fragmented

A Skill may work in Work or desktop but not in a normal web or phone chat. A plugin may have different supported surfaces from a standalone Skill. Workspace administrators can also restrict Skill and plugin access.

State the supported surfaces in any guide or share page. Do not make people install something before telling them where it runs.

Interfaces and availability will change

Skills are new enough that buttons, menus and supported surfaces may move. Add a “last reviewed” date to published tutorials and link to OpenAI's current documentation.

📝How to install and use an existing Skill

The exact interface depends on the surface and account, but the basic flow is:

Open the Skill's shared or directory page. - https://chatgpt.com/skills
Install or enable it.
Open a supported ChatGPT Work or desktop conversation.
Type @ and select the Skill.
Describe the result you want and provide the required inputs.

For example

@Meal Planner Turn the ingredients I already have into a simple dinner plan. Prioritise anything close to expiring, respect my dietary preferences, and don’t invent ingredients I haven’t mentioned.

You do not need to repeat the entire workflow. That is the Skill's job. You still need to provide task-specific context: the notes, audience, deadline, requested format and anything unusual about this run.

Explicit and automatic activation

ChatGPT can activate a Skill in two ways:

Explicitly: you select it with @.
Implicitly: ChatGPT decides the request matches the Skill's description.

Explicit selection is better when using the wrong workflow would waste time or produce the wrong output.

🛠️ How to create your first Skill

Open ChatGPT Work and invoke @skill-creator.

Start with a concrete example instead of describing an imaginary universal assistant.

@skill-creator Help me create a Skill (add the skill describtion you want to create)

The creator should ask what the Skill does, what requests should activate it, what the finished result looks like, and whether it needs reusable files or scripts.

Step 1: Define one job (this is just examples you can work with your AI on this)

Finish this sentence:

When I say [type of request], the Skill should [produce this result using this method].

Weak:

Help me with marketing.

Stronger:

Turn one long article into five platform-specific social posts, preserving its factual claims and changing the hook, length and call to action for each platform.

Step 2: Give it real examples

Useful examples reveal details that abstract instructions miss.

Provide:

one or two realistic requests;
a good previous result, if available;
a template or reference file;
recurring mistakes it must avoid;
the audience and format;
the point where it should stop and ask for clarification.

Step 3: Decide what belongs inside

Use the smallest structure that does the job.

Put the core workflow in SKILL.md.
Put long guidance and examples in references.
Put templates, images and brand assets in assets.
Add scripts only when deterministic execution is genuinely useful.

Do not add files merely because folders exist. A Skill is not a digital junk drawer.

Step 4: Create and validate it

Ask the creator to build, validate and install the Skill.

Create the Skill, validate its structure, install it, and tell me what prompt to use for the first test.

Validation checks the package structure and metadata. It does not prove the workflow is good.

Step 5: Test it with a real task

Use the Skill as you normally would. Do not give it a suspiciously perfect example designed to flatter the instructions.

Check:

Did it activate?
Did it ask for missing information?
Did it follow the required process?
Did it preserve facts and constraints?
Did it produce the correct format?
Did it use the reference files properly?
Did it inspect or validate its own result?

Fix the smallest instruction responsible for the failure. Do not respond to one bad output by adding six pages of panic legislation.

📂 Adding files, templates and images

Skills become more useful when they carry the materials needed for the job.

Templates

You can bundle document layouts, presentation templates, spreadsheets, boilerplate projects or plain-text structures. Tell the Skill when to copy, fill or preserve each template.

Reference documents

Reference files can hold policies, terminology, brand rules, technical schemas or detailed examples. Keep them separate from the main workflow so ChatGPT only loads them when needed.

Visual references

Image assets can preserve a recurring character, mascot, product, logo or visual style.

Separate their jobs:

identity references establish appearance;
full-body references establish proportions;
style references establish rendering, lighting and palette;
composition references establish framing or staging;
“avoid” references show recurring failures.

One image should not silently define every category.

Scripts

Scripts suit jobs that must perform the same mechanical operation reliably, such as transforming files or validating a known format.

They also increase maintenance and environment requirements. Do not add code to a Skill whose whole job can be expressed clearly in twenty lines of instructions.

📦 Worked example: a reusable visual identity

We created a generic Build Visual Identity Skill to help people organise reference-led image generation.

The public version contains blank identity profiles, style templates, reference-image folders and a quality checklist. It contains no private identity photographs.

A person installs the public template, uploads their own references and asks ChatGPT to create a private personalised copy:

@skill-creator Use @Build Visual Identity as the template to create and install my private image-generation Skill. Ask me for the necessary references and appearance details.

The resulting private Skill can contain:

curated face and body references;
fixed appearance traits;
flexible styling traits;
scale rules between recurring subjects;
preferred rendering and lighting;
accessories and continuity details;
common generation failures;
a post-generation quality checklist.

It improves consistency because the reference system and checking process travel with the workflow. It still does not train a new image model, and it cannot guarantee perfect likeness.

👷‍♀️ How to share a Skill safely

Before sharing, decide whether people should use your exact workflow or create personalised copies.

Share the exact workflow

Suitable for public standards, checklists, generic templates and workflows containing no confidential material.

Share a setup template

Suitable for writing voices, personal identity references, company branding or any workflow that needs private files. Share empty profiles and instructions; let each person add their own material privately.

Package it as a plugin

Skills are the workflow format. Plugins are the distribution format when you want people to install a reusable bundle, particularly when it includes connected tools or several related Skills.

Public plugin distribution may require packaging, testing, metadata, privacy information and review. A local or personal Skill working for you does not prove a public plugin will be approved.

☢️ Privacy checklist

Before installing or sharing a Skill, ask:

What files are bundled inside it?
Does it contain personal photographs or confidential examples?
Does it connect to external services?
What permissions will those services request?
Is the Skill private, shared with a workspace or publicly distributed?
Can private material be replaced with placeholders?
Does the workflow tell ChatGPT to expose source material in its output?

Never bundle credentials. Never assume “private” means “safe to fill with secrets.” Follow your account and organisation's data rules.

⚠️ Troubleshooting

The Skill does not appear

Confirm it is installed and enabled.
Check that you are using a supported surface.
In ChatGPT, open the Skills section under Plugins and refresh or reopen it if the list is stale.
Start a new supported Work conversation and type @ again.
Check whether a workspace administrator has restricted Skills.

ChatGPT ignores the Skill

Invoke it explicitly with @Skill Name.
Make the request match the Skill's stated purpose.
Rewrite an unclear Skill description so its main job appears early.

The result follows the format but misses the point

The Skill may describe output structure without teaching the decision process. Add one realistic example and clarify how it should judge trade-offs.

The Skill keeps asking unnecessary questions

State which assumptions it may make and which missing details genuinely require confirmation.

The Skill invents facts

Add an explicit evidence rule: distinguish supplied facts from inference, label uncertainty and leave unknown fields blank rather than guessing.

Reference images are copied too literally

Label each reference by role. Tell the Skill what evidence to borrow and what temporary details not to inherit.

The Skill became worse after an update

Remove instructions added in reaction to isolated failures. Look for contradictions, duplicated rules and long passages that bury the core workflow.

🥰 A practical first Skill

Choose something you already repeat and already know how to judge.

Good first candidates:

a weekly update;
a recurring article structure;
meeting follow-up notes;
a document review checklist;
a social-content adaptation workflow;
a visual-reference organiser.

Give @skill-creator one real example, explain what a good result contains, and list the mistakes that matter. Test the first version on ordinary work. Then improve it when evidence gives you a reason.

Quick-start prompts

Create a simple instruction-only Skill

@skill-creator Create a Skill that [specific job]. It should trigger when [example requests]. A good result must include [requirements] and avoid [failures]. Keep it instruction-only unless a reusable file is genuinely necessary.

Create a Skill from an existing template

@skill-creator Create a private Skill using the attached template. Preserve its structure, replace the placeholders with my information, validate it, install it, and give me a realistic test prompt.

Improve an existing Skill

@skill-creator Update my [Skill name] Skill. Here is the request that failed and the result it produced. Identify the smallest reusable instruction change that would prevent this failure without overfitting to this one example.

Build a visual identity Skill

@skill-creator Use @Build Visual Identity as the template to create and install my private image-generation Skill. Ask me for identity, full-body and style references, separate their roles, add a quality checklist, and do not treat the references as model training.

Official references

With Love

Firecracker&Cass

build-visual-identity.zipZIP

Download file