GPT 6 skills

GPT-6 model compatibility

The original examples were run with Astra. New September 24 checks apply frontend-release-check with Sol to a code fixture and csv-quality-triage with Luna to eight support records. Each record identifies the steps exercised and the parts that need a different input or browser. Use the Model filter above to find those tested resources.

What GPT 6 skills add to a prompt

A skill turns a repeated task into a named workflow: when to use it, which inputs to collect, what steps to follow, and how to check the result. A prompt is convenient for one request. A skill keeps the procedure available for later requests without pasting it again. Start with the prompt library when you need a single instruction; use a skill when the same checks matter across several tasks.

“GPT 6 skills” describes workflows you can use with GPT 6 in a skill-capable host such as Codex. The model handles the reasoning; the host supplies files, tools and permissions. Installing a skill does not add model access, an API subscription or a browser integration. A writing checklist can work with text alone. A browser-testing skill needs an actual browser tool.

The Agent Skills format uses a folder containing SKILL.md. Its YAML frontmatter names the skill and describes when it applies; the Markdown body contains the procedure. Extra references or scripts can accompany it. The description helps the host discover the right workflow, while the body supplies detailed instructions when needed. A plugin can distribute skills alongside tools and integrations; a single skill does not need all of those features.

Install a skill in Codex

For a first try, choose one of the original downloads above. Each ZIP contains one named folder. Extract that folder and keep its supporting files together. In a repository, the resulting path should look like this:

your-project/
  .agents/
    skills/
      meeting-action-ledger/
        SKILL.md
        example-input.md
        first-output.md
        check-result.json
        LICENSE

Codex discovers repository skills in .agents/skills. For personal skills across projects, use ~/.agents/skills instead. Keep the folder name identical to the name in the frontmatter. Avoid nesting the extracted folder twice: Codex should reach meeting-action-ledger/SKILL.md, not another same-named directory inside it.

In Codex CLI or the IDE extension, use /skills to inspect available skills or mention $meeting-action-ledger in your request. Provide the meeting notes and ask for the action ledger. If the new skill does not appear, restart Codex and check the folder path. The current OpenAI guide covers supported hosts and discovery behavior. For a plugin-bundled recommendation, use that publisher’s installation instructions so its companion tools are included.

Copy SKILL.md copies the entire file, including frontmatter. Copy example input copies the complete skill followed by its sample task; it never includes the model’s answer. You can paste that version for a direct trial before installing anything. For community skills, open the upstream folder and read both the skill and any referenced scripts. Use the author’s package rather than copying one file while leaving required references behind.

Choose a workflow that fits your task

Pick a specific job before choosing a popular repository. Systematic Debugging is useful when you can reproduce a bug and run the project’s tests. React guidance fits a React codebase; it adds little to a plain document task. Database guidance needs the real schema and workload. A design-search tool can suggest a direction, but the final page still needs inspection at the intended screen sizes.

This directory draws from 23 researched candidates and links to 12 selected upstream skills. Repository activity and adoption help us find candidates; they do not measure the quality of an individual skill or predict its performance with GPT 6. We favor clear workflows, identifiable authors, concrete source files, and setup requirements a reader can understand. “Official” means distributed through OpenAI’s current repository. “Community” includes independent maintainers and other vendors; their original authors remain credited.

OpenAI’s former openai/skills repository now points readers to openai/plugins. The official entries here use the current locations. Third-party source code is linked, not repackaged. Each publisher’s license and dependencies still apply. Our six original packages are MIT-licensed and contain instructions and small example files, with no bundled executable installer or external service requirement.

How the original skills were checked

All six example tasks were run in Codex with GPT-6 Astra at medium reasoning. Each skill was explicitly read and used in one shared session. The downloadable inputs are synthetic; the output shown with each package is the recorded model response. These runs check the examples, not automatic skill selection or every possible input. The recommended upstream skills were reviewed from source; they were not included in this execution set.

The checks focus on observable details. The meeting example keeps a proposal separate from approved decisions and leaves an unnamed owner unassigned. The research example preserves a conflict between two seat limits instead of inventing a resolution. The document comparison covers six changed fields and two unchanged controls. The CSV example handles a quoted comma, duplicate records, missing amounts and separate currency totals.

The frontend example is an offline review of a small HTML fixture. It identifies source-level issues; it does not claim a browser accessibility or performance test. For an actual browser exercise and its repair record, see frontend QA. The content-refresh example preserves existing URLs and headings while linking proposed changes to supplied evidence. The practical guide explains how to carry the same input–output–check discipline into a larger task.

Adapt a skill without losing its checks

Change the inputs and output format first. Keep the parts that protect your actual workflow: source locations for research, separate currency totals for data, missing owners for action lists, and explicit stops before publication. Replace generic labels with the documents or project conventions you really use.

Try one ordinary case and one difficult case after editing. A meeting with no deadlines, a CSV with an empty field, or two sources that disagree can expose an assumption quickly. Compare the new output with your checklist; a polished answer can still omit a required item. Save the input and the result together so the next revision has a concrete comparison. When your procedure stabilizes, give it a specific description that explains when it should be selected.