Terms in this guide: AGENTS.md · Context
Choose the right Codex building block by deciding whether you need a one-time request, durable project context, a reusable procedure, connected tools, or an installable bundle.
Start with the job that needs to persist
The labels become clearer when you ask what must survive the current conversation. A prompt expresses the immediate result you want. AGENTS.md gives Codex durable project guidance. A skill packages a reusable method. MCP connects tools or external context. A plugin is an installable bundle that can bring skills, MCP servers, browser capabilities, and lifecycle hooks together.
These pieces often cooperate. A request can activate a skill, the skill can follow repository instructions, and an MCP tool can retrieve the evidence needed to finish the work. Choosing one does not mean excluding the others; it means assigning each responsibility to the narrowest layer that can own it reliably.
Use prompts for the current outcome
A good prompt identifies the task, important context, constraints, and the evidence that will make the result reviewable. It can be detailed, but it should remain specific to the work in front of you. If you keep pasting the same procedure into unrelated conversations, that procedure is a candidate for a skill.
Keep project facts out of generic reusable prompts when those facts already belong in the repository. This reduces contradictory instructions and makes the request easier to adapt.
Use AGENTS.md for repository expectations
AGENTS.md is the place for durable facts and expectations that should apply whenever Codex works in a repository or a scoped directory: validation commands, ownership boundaries, formatting conventions, or where the application starts. Nested instructions can narrow the guidance for a component.
Do not turn AGENTS.md into a catalog of every possible workflow. A long file full of optional procedures competes with the facts every task needs. Move a repeatable specialist procedure into a skill and keep the repository file focused on the project.
Use a skill for a repeatable method
A skill is a directory with a required SKILL.md and optional scripts, references, assets, or agent metadata. Its description matters because Codex can select a skill when the request matches that description. Write the trigger around a recognizable user task and name important boundaries.
Use a skill for a review method, publishing workflow, evidence checklist, or transformation that benefits from the same decision process each time. Inspect downloaded files before use, and distinguish a structurally valid package from one evaluated on real tasks.
Use MCP and plugins for connected capabilities
MCP servers expose tools and context. They may require authentication and can perform actions in another system, so evaluate permissions, data handling, and the exact operations each tool provides. The local Codex clients can share MCP configuration on the same host.
Plugins make capabilities easier to install and share. They can include skills and MCP servers, while supported interfaces vary. Check current product documentation before writing setup instructions that assume the same plugin behavior in Desktop, CLI, IDE, web, or mobile.
A practical decision rule
Ask five questions: Is this instruction only for the current result? Is it a durable fact about this repository? Is it a method worth reusing? Does it require a connected tool? Should several capabilities be distributed together? The answers map naturally to prompt, AGENTS.md, skill, MCP, and plugin.
Continue with the installation guide if you want to use a downloaded skill, or the skill-building tutorial if you want to turn your own procedure into a package.
Primary source
OpenAI skills and plugins documentation. Source checked September 10, 2026.
Progress stays in this browser. No account needed.
Updates and upgrade guidance
- Learning context added
Added prerequisites, a suggested practice estimate, expected output, and connected definitions. The article procedure was not independently re-evaluated in this change.
What to do: Use the opening checklist to prepare. No software update is required for this editorial change.
This log starts with the first recorded review; it is not a reconstruction of earlier revisions.