In brief: A skill is a reusable workflow with instructions, examples, and, if needed, scripts. A plugin is an installable package that can combine skills, connections to external services, MCP tools, and other components. For a recurring task within Codex, a skill is usually enough; working with GitHub, Google Drive, Slack, a CRM, or another external system more often requires a plugin.
Skills and Plugins help turn Codex from a general-purpose assistant into a working tool for a specific process: preparing reports, reviewing documents, handling development tasks, or collecting data from connected services. But the choice should start not with a list of trendy extensions, but with a recurring operation that can be described, verified, and measured.
How a Skill differs from a Plugin
In current OpenAI terminology, a skill is a set of instructions and supporting resources for a specific task or workflow. A plugin is an installable package that can contain one or more skills, connectors for external services, MCP servers, and additional components.
| Criterion | Skill | Plugin |
|---|---|---|
| Primary role | Defines a repeatable way to complete one clear task | Packages a broader scenario and the required connections |
| What it can include | Instructions, templates, examples, reference materials, scripts | Skills, connectors, MCP tools, settings, and other components |
| Access to external data | Not required; depends on available tools | Can connect external services and their data |
| Typical example | Prepare a weekly report using a specified template | Retrieve data from Google Drive, process it, and send the result to Slack |
| When to choose it | The process is narrow, recurring, and does not require a separate integration | External services, several related capabilities, or distribution of a ready-made package are needed |
A simple rule of thumb: if the task sounds like “always do this following these steps and in this format,” start with a skill. If it sounds like “get data from here, perform actions there, and follow an overall workflow,” consider a plugin.
How a Skill works in Codex
A skill is stored in a separate folder. The required file SKILL.md contains the name, scope description, and instructions. Scripts, reference materials, and templates can be stored alongside it. Codex first sees the skill's brief description and loads the full instruction file when a task matches its purpose. This approach saves context and makes a large set of specialized workflows available.
my-skill/
├── SKILL.md
├── scripts/
├── references/
└── assets/The most important part is not the length of the instructions, but a clear scope. The description should specify what task the skill solves, when it should be triggered, and when it should not be used. A skill that is too general starts triggering at inappropriate times, while one that is too narrow must be invoked manually all the time.
Explicit and implicit invocation
In Codex, a skill can be selected explicitly using a mention through $. Codex CLI and the IDE extension also provide a list of skills. The second option is to describe the task in plain language: Codex will match the request to the purpose of installed skills and automatically select the appropriate workflow.
Explicit invocation is useful when the result must strictly follow a specific process. Implicit invocation is better when it is more important to set the task quickly and leave the tool choice to the agent. For critical processes, it is useful to specify the skill explicitly and save test examples of the expected result.
What a Plugin includes
A plugin can include not only instructions, but also access to work systems. For example, a connector allows Codex to work with GitHub, Google Drive, or Slack within the permissions granted to the user. An MCP server provides structured tools and context, while the skill explains to the agent in what sequence to use these capabilities.
As a result, a plugin solves not an individual text task, but an entire operational scenario. For example:
- find requirements in Google Drive, check the implementation in the repository, and prepare a report;
- review pull request comments, make the approved changes, and recheck CI;
- collect data from a spreadsheet, create a presentation, and save the final file;
- read new requests, classify them, and prepare draft responses without sending them automatically.
This is close to the logic of AI agents and assistants: value is created not by an individual model response, but by a controlled sequence of actions, access, and checks.
Where Plugins and Skills are available
As of August 2026, plugins are available in ChatGPT Work on the web, in ChatGPT Work or Codex in the ChatGPT desktop app, and through the plugin browser in Codex CLI. Plugins are not supported in regular Chat mode, the mobile app, or the Codex IDE extension. Standalone skills can be used in the ChatGPT desktop app, Codex CLI, and the IDE extension.
The availability of a specific plugin depends on the environment, workspace settings, and authentication method. Some connections require OAuth and a separate sign-in to the external service. Therefore, a plugin being listed in the catalog does not mean it automatically has access to data.
How to install and use a Plugin
In ChatGPT Work and the desktop app
- Open Work or Codex mode on a supported surface.
- Go to the Plugins section.
- Find the required plugin, open its card, and click the plus button.
- If a connector is required, sign in and review the requested permissions.
- After installation, start a new chat and describe the result you want to get.
A plugin can be selected explicitly through @ or simply formulate the task. For example: “Gather the latest requirements from Google Drive and prepare a checklist for the team.” In the second case, the system selects the appropriate installed capabilities itself.
In Codex CLI
In Codex CLI, the catalog opens with the command /plugins. After installing a plugin, you need to start a new session for its skills and tools to become available. In the catalog, you can browse plugin sources, check installed packages, and enable or disable them.
How to create your own Skill
The best candidate for your first skill is a process the team already performs regularly and explains from scratch every time. This could be reviewing an article before publication, preparing a sales proposal, investigating an incident, updating documentation, or producing a weekly report.
1. Define the task boundaries
Describe one specific outcome. Instead of “help with marketing,” it is better to say: “based on the incoming brief, prepare an article plan, verify the required facts, and produce HTML using the editorial template.”
2. Define the input and final result
- what files, links, or data the skill receives;
- which steps are required;
- what format the output should have;
- which actions are prohibited without confirmation;
- how to verify the quality of the result.
3. Create a draft
In Codex, you can invoke the built-in $skill-creator and describe the process in plain language. Alternatively, manually create a folder and file SKILL.md. For the first prototype, instructions are usually enough; add scripts only where they truly improve repeatability or enable validation.
4. Test with real examples
Use several typical tasks and at least one awkward edge case. Check not only for a polished result, but also for skipped steps, made-up data, incorrect file operations, and attempts to perform a critical operation without confirmation.
5. Improve using the error log
A good skill does not come from a long, generic prompt, but from observing the workflow. If the team constantly fixes the same format, adds the same check, or sends the result back to the agent for revision, that is a signal to clarify the instructions or add a test.
Practical business use cases
Content and editing
A skill can store the tone of voice, article structure, linking rules, fact-checking requirements, and a final checklist. A plugin is needed if the workflow must find materials in Google Drive, receive tasks from a project management system, or publish results in an external environment.
Sales and support
A skill defines request classification rules, response structure, and conditions for handing off to a manager. A plugin connects email, CRM, or a messenger. In sensitive scenarios, it is better to assign AI the preparation of a draft and data collection, while leaving message sending, deal status changes, or refunds to a person.
Development
A skill can describe code review rules, testing standards, and the report format. A plugin adds access to the repository, pull requests, tasks, and CI. Such a workflow is useful when it not only writes code, but also checks changes in a controlled environment and shows the developer a clear diff.
Operational reports
For a single local file, a skill with calculation and formatting rules is enough. If the data is in Google Sheets, a CRM, and several reports, a plugin helps bring the sources into one process. Similar setups relate to AI business process automation, where repeatability, action logs, and exception control are especially important.
Security: what to check before connecting
A plugin receives capabilities only within connected services and granted permissions, but those permissions may include reading and changing work data. Before installing, assess not only the benefits, but also the consequences of an error.
- Least privilege. Grant access only to the data and actions required by the scenario.
- Separate read and write access. For the first pilot, keep the workflow in analysis and draft mode where possible.
- Confirm critical actions. Sending messages, deleting data, changing permissions, and publishing should have a clear control point.
- Verify the source. Review the plugin listing, package contents, connected services, and authorization terms.
- Test environment. Test a new workflow on copies of data or a limited set of objects.
- Logging and reversibility. Actions must be visible, and an incorrect result must be fixable.
If a plugin uses hooks—commands run at specific points in the lifecycle—they should be reviewed separately before enabling. For processes involving personal or commercially sensitive data, you must also check internal company policies and applicable data protection requirements.
How to launch a realistic MVP
Do not start by trying to automate an entire department. Choose one frequent process with clear input and verifiable output. For the first cycle, the following approach is enough:
- document the current steps and manual checks;
- choose one recurring task;
- build a skill without unnecessary integrations;
- test it on real examples;
- connect a plugin only for the systems the workflow cannot work without;
- keep human confirmation for important actions;
- compare the result with the original process.
Four basic metrics are suitable for evaluating an MVP: task completion time, the share of results that need no manual rework, the number of errors, and the number of actions that still require a person. For customer-facing processes, add first response speed and the share of requests correctly handed off to a manager. These indicators are more useful than the general impression that “AI has started working better.”
Common mistakes
- Creating one universal skill for dozens of unrelated tasks.
- Installing plugins before the business process itself is understood.
- Granting write permissions when read access is enough for the pilot.
- Expecting that connecting a service will automatically fix imprecise instructions.
- Not testing results on real and edge-case examples.
- Measuring success by the number of installed tools rather than the quality and time required to complete work.
FAQ
Can you use a Skill without a Plugin?
Yes. If a task is based on instructions, local files, and already available tools, a separate plugin may not be needed. A plugin is primarily needed to install a ready-made capability package or connect external services.
Do you need to explicitly invoke a Skill every time?
No. Codex can select a skill automatically if the request matches its description. Explicit invocation through $ is useful when you need to ensure a specific workflow is applied.
Why is an installed Plugin not visible in the current chat?
After installation, you usually need to start a new chat, or a new session in Codex CLI. Also check whether the selected surface supports plugins and whether the external service connection is complete.
Can I trust a Plugin to automatically send emails or update the CRM?
Technically, it depends on the tools and permissions, but at the MVP stage it is safer to prepare drafts and request confirmation before external or hard-to-reverse actions. Automation can be expanded after testing, permission setup, and error review.
When should you turn a custom Skill into a Plugin?
When the workflow needs to be distributed as an installable package, combined with other skills, or extended with connectors and MCP tools. For a local process in one project, a separate skill folder is usually simpler.
What to do next
Start with one task the team repeats every week. Describe the input data, required steps, result format, and manual review points. If external systems are not needed yet, build a skill. If the process requires data or actions in other services, add a suitable plugin and the minimum required permissions.
Need help designing such a workflow for your team? Discuss the task with CenterAI: we will define the right MVP, required connections, control points, and outcome metrics.
Sources
Last checked: 09.08.2026.