Most of us are becoming familiar with prompting generative AI tools such as ChatGPT or Claude. We ask a question, provide some context, refine the response and hopefully arrive at something useful.
But what if, rather than explaining how we want an AI to approach a task every time, we could give it a reusable set of instructions, examples and resources? That's the idea behind Agent Skills.
An Agent Skill is essentially a reusable package that helps an AI assistant carry out a particular type of task.
Rather than relying on a single prompt, a skill might contain:
Instructions describing how to approach the task.
Examples of good outputs.
Templates, reference materials or design standards.
A sequence of steps to follow.
Scripts or tools for completing more technical work.
Many skills use a file called SKILL.md, which tells a compatible AI agent what the skill does and how to use it. Related approaches use files such as AGENTS.md to provide instructions for agents working within a particular project.
The important distinction is that we're not necessarily making the AI more intelligent. We're giving it a better process to follow.
For those of us working in universities, that could mean more consistent teaching resources, structured research workflows, improved document production or more reliable administrative processes.
https://github.com/zarazhangrui/beautiful-html-templates
This was the example that first sparked my interest (thanks to Dr Ian Stonebridge for sharing it).
This is a GitHub repository called Beautiful HTML Templates, and wihle GitHub can be a bit daunting for many, myself included, it has a surprisingly simple instruction which is to paste this sentence into your AI tool of choice:
"Clone https://github.com/zarazhangrui/beautiful-html-templates and follow the instructions in AGENTS.md to build me a beautiful HTML slide deck."
The repository contains a collection of HTML presentation templates, along with instructions explaining how an AI coding agent should use them.
Rather than immediately generating a presentation from scratch, the agent is guided through a process:
Understand the presentation's purpose, audience and preferred visual style.
Explore the available design templates.
Select three promising approaches and produce examples.
Invite the user to choose a preferred design.
Build the complete presentation using the selected style.
Preview and refine the finished output.
The interesting thing isn't simply that the templates look attractive. It's that the AI is being given a design process to follow, including opportunities for human judgement.
This could be particularly useful for developing teaching presentations, online learning materials or interactive resources.
For example, I've spent considerable time developing HTML pages for our Canvas virtual learning environment. A similar approach could potentially help maintain consistent standards for accessibility, navigation and learning design while allowing individual lecturers considerable creative freedom.
One important distinction: this example uses an AGENTS.md file rather than a conventional SKILL.md package. It's closely related to the Agent Skills approach, but the two formats aren't identical.
GitHub has become an interesting place to explore these resources. Here are several collections that I've found useful.
https://github.com/topics/agent-skills
A broad starting point for exploring repositories tagged with Agent Skills.
This is particularly useful if you're curious about what's available rather than looking for a specific application.
You'll find resources relating to coding, design, research, productivity and automation. The quality and complexity vary considerably, but it's a good place to browse.
https://github.com/neverbiasu/awesome-research-skills
This is probably the most immediately relevant collection for academics.
It brings together more than 150 research-oriented skills covering the research process, from developing ideas and reviewing literature through to analysis, visualisation, academic writing and peer review.
Some particularly interesting sections include:
Qualitative and mixed-methods research
https://github.com/neverbiasu/awesome-research-skills#qualitative--mixed-methods
Includes skills for thematic analysis, grounded theory, qualitative coding, analytic memo writing and mixed-methods approaches.
Statistical analysis
https://github.com/neverbiasu/awesome-research-skills#statistical-analysis
Includes skills for selecting statistical approaches, checking assumptions, conducting analyses and producing research outputs.
Literature reviews
https://github.com/neverbiasu/awesome-research-skills#literature-review
Includes structured approaches to literature searching, critical reviews, research gaps and citation management.
These could be useful for researchers, supervisors and postgraduate students, although the quality of any analysis still needs careful evaluation.
https://github.com/alfarius42/awesome-agent-skills
A broader collection featuring skills published by development teams and independent contributors.
Much of the material is technically oriented, but there are potentially useful resources for creating websites, documents, visual materials and more complex workflows.
This is worth exploring if you're interested in building digital teaching resources or experimenting with AI-assisted development.
https://github.com/anthropics/skills
Anthropic's public collection provides examples of skills and how they are organised.
A useful place to start if you want to understand the structure behind Agent Skills rather than simply browsing examples.
https://github.com/openai/skills
OpenAI also maintains a collection of skills for coding-agent workflows.
These provide another opportunity to explore how reusable instructions and resources can be packaged to support different tasks.
There are several areas where I can see potential value.
Teaching and learning: Skills could support the creation of accessible learning materials, interactive activities, assessment guidance and consistent VLE pages.
Research: Structured workflows could help with literature exploration, methodological planning, data analysis, visualisation and manuscript preparation.
Student recruitment: A skill could combine approved programme information, audience needs and communication guidance to support the development of recruitment materials.
Professional services: Reusable workflows could support reporting, documentation, communications and other recurring administrative tasks.
Perhaps the most interesting possibility is developing our own skills that reflect institutional practices and expectations.
For example, rather than asking lecturers to master increasingly complicated prompts, a university might provide a reusable learning-design skill incorporating accessibility requirements, curriculum principles, templates and examples.
That could help establish shared standards without requiring every lecturer to produce identical materials.
Not every Agent Skill is safe, reliable or suitable for institutional use.
Some are simply text-based instructions. Others include executable scripts, connect to external services or request access to files and accounts.
Before using an unfamiliar skill, it's worth considering:
Who created it, and can its instructions and code be inspected?
Does it need access to sensitive information or university systems?
Does it execute code or communicate with external services?
Is it compatible with the AI tool you're using?
Does it align with institutional information governance and AI policies?
For research, there's an additional concern: a skill that produces an apparently sophisticated analysis isn't necessarily following an appropriate methodology. Researchers remain responsible for methodological decisions, interpretation and verification.
A useful starting point is to experiment with publicly available, non-sensitive material before considering more consequential applications.
What interests me most about Agent Skills is the shift from asking AI to produce something towards giving AI a structured process for producing it.
This doesn't remove the need for human expertise. In many cases, it makes that expertise more important: someone still needs to decide what good practice looks like, what the workflow should involve and whether the resulting output is any good.
I'm particularly interested in whether these approaches could support more consistent, accessible and engaging digital learning materials, while retaining the creativity and agency of individual lecturers.
For now, I'm treating Agent Skills as something worth experimenting with rather than another technology everyone needs to adopt.
The HTML slideset example seems like a good place to start.
BJ 18.11.26