Find, create and use AI skills: a practical trial guide

Know what you are choosing

A prompt describes the current request. A skill packages a reusable process. A connector supplies access to tools or data. A plugin can bundle capabilities for installation. A skill can guide a connected action, but it does not supply credentials or run on a schedule by itself.

The Agent Skills format centers on SKILL.md, with a name, a description and instructions. Optional references, assets and scripts support the job. Check the host’s compatibility and dependencies as well as the file format.

Where to find existing skills

Start with the skills available in your app, then inspect official publisher catalogs. Search by the result you need rather than a broad label such as “productivity.”

• OpenAI skills catalog: examples and curated skills for Codex. Use the installation route described in its current documentation.
• Anthropic skills repository: examples, templates and Skill Creator. Check the instructions and license for the particular package.
• Your product’s directory: look for skills or plugins available to your account and workspace. An installed plugin may include both skills and connected tools.
• Your own process: a good completed example and a checklist can become the brief for a custom skill.

Before trying one, record its publisher, source, last relevant update, intended host, dependencies and requested access. A downloaded file is not necessarily an installed capability. Do not assume a skill from another host will work unchanged.

A five-step skill trial

Use the Tool Trial Lab to keep your comparison concrete. Write the input once and judge each result against the same criteria.

When to use Skill Creator

Create a skill when your preferred method is clear but no existing option follows it. Capture a process you perform repeatedly and can evaluate yourself.

1. Show the creator one good input and output.
2. Describe when the method should and should not apply.
3. Ask for concise instructions with a clear output format.
4. Test an ordinary request and two edge cases.
5. Save or install through your host and select it for a fresh task.

In ChatGPT Work, invoke @skill-creator; in Codex, use $skill-creator. Anthropic publishes a separate Skill Creator implementation for supported Claude environments. Follow the installed package’s invocation instructions.

Create an instruction-only skill named tool-trial-summary.
Use it when I provide two tool trial records for the same job.
Compare correctness, missing facts, review time and required access.
Return a decision with evidence and remaining unknowns.
Do not infer a winner from popularity or invented scores.
Ask for missing trial data. Test a tie, an incomplete record
and a result that is fast but inaccurate. Help me save the skill.

The illustrated AI Teachers walkthrough covers creation, testing and saving.

Worked example: meeting follow-up

Imagine two candidates receive these fictional notes: “Maya will send an outline Friday. A designer was suggested. Budget not agreed.” Candidate A drafts a polished email that assigns the designer and invents a budget. Candidate B leaves those points unresolved. Candidate B preserves the evidence better, even if you still edit its prose.

For the second run, remove the deadline. For the third, add conflicting deadlines. A useful result identifies the gap or conflict. Record what actually happened; do not mark a case as passed because the skill promises to handle it.

Advanced use: references, tools and handoffs

Keep the recurring method in the main instructions and current facts in maintained references. Add templates for stable output formats. Add scripts for exact operations you can test with known inputs.

A connected workflow needs a clear source and authorized tool access. If research, drafting and review use separate skills, specify what passes between them: sources, dates, claims, open questions and the deliverable. Give the reviewer the original evidence. A second assistant’s agreement is not independent verification.

Compare with a baseline, including human correction time. Keep unseen test examples for revisions. Assign an owner and retain a known working version. Sharing and scheduling need separate setup; copying a skill folder does not transfer a signed-in account.

Make one useful choice today

Choose one recurring task, inspect one skill and run one example. Keep it if it improves the result after review; adapt it if the method nearly fits; remove it if setup outweighs the benefit. Continue with the tool test-bench guide or RAG, MCP and agents explainer.