At the HELish Summit, Antti Pajunen asked me about the skills I keep recommending. I have mentioned them at a lot of events, but I have never put them all in one place. These are the ones I return to when I work with AI.

Where I find skills

Skills.sh

Skills.sh is where I find skills. You can browse what people are building, and it shows security audits too.

A skill is a set of instructions for an AI agent that can often read files, run commands and call tools on your behalf. Look at the available audits and read the skill itself before you install it. A useful skill can still be wrong for your environment.

Microsoft's agent skills gallery

Microsoft's agent skills gallery is made by the Microsoft CAT team. It is worth a look if you use Copilot Cowork, Copilot Studio or Scout. You can filter by platform and tag, and the gallery takes community contributions too, so you can share work that others might find useful.

Each entry includes its instructions and downloads, so you can see what you are adding before you use it. Still do your own review, but it is a useful starting point for Microsoft-focused work.

Skills I keep using

Each skill below includes an example you can click to expand.

Grill Me

Matt Pocock has a bunch of great skills. Have a look through his collection. The one I use most is Grill Me. It asks me the right questions after I have made a plan in plan mode and before I implement it. That gives me a useful sequence: plan mode to shape the approach, Grill Me to stress-test it, then implementation. I would rather get that pushback while the approach is cheap to change than after I have built the wrong thing efficiently.

Example: review a plan with Grill Me before implementing

Goal

Add a command that imports a CSV file, validates each row, and shows a summary of the rejected records.

Prompt

I created the plan above in plan mode. Before I implement it, grill me on the plan in rounds. Ask the whole frontier of questions for each round, number them, and give your recommended answer for each one. Start with assumptions and the user workflow, then use my answers to challenge the scope, edge cases, error handling, and success criteria. Wait for my answers, point out anything the plan needs to change, and only stop when we have reached a shared understanding and it is ready to implement.

Outcome

The first round might look like this:

❓ Q1 - Who can import a file?: Who is allowed to upload the CSV, and what should happen if two people import the same file?

➡️ The command should be available to authenticated users, and duplicate imports should be detected and reported rather than processed twice.

---

❓ Q2 - What makes a row valid?: What validation rules should apply, and should validation stop at the first error or report every problem?

➡️ Validate every row and report all problems so users can fix the complete file in one pass.

The next round would use my answers to ask deeper questions about error handling, the rejected-record summary, and edge cases. Answering the rounds helps identify changes to make to the plan and decide when it is ready to implement.

Frontend Design

Frontend Design by Anthropic is another great one. It helps a lot, though it is not a 100% solution. Use it as a starting point and edit it to cover what you need.

For me, that meant adding guidance for HTML and CSS design for Outlook. I use it often with Copilot Cowork, and I do not want it producing CSS that Outlook does not support. That addition makes it much more useful for my work.

Example: use Frontend Design for an Outlook-safe email

Goal

Create a responsive welcome email for a new event attendee.

Prompt

Use a clear primary action, keep the HTML and CSS compatible with Outlook, and explain which layout choices are fallbacks for email clients with limited support.

Outcome

A welcome email design with a clear primary action, Outlook-compatible markup, and documented fallbacks for email clients with limited support.

Impeccable

Talking about design, Impeccable is one of my favourites. It is a design plugin for apps, websites and more. I used it for this site too.

I find it most useful when something works but still feels like a first version. It helps me turn a functional idea into something I want people to use.

Example: use Impeccable to improve a working page

Goal

Improve an event registration page without changing its existing functionality.

Prompt

Review this page as a product designer. Keep the existing functionality, but improve the visual hierarchy, mobile layout, accessibility, and empty or error states. Make the recommendations concrete enough to implement in the current codebase.

Outcome

Concrete design recommendations that make the working page feel more polished and are ready to implement in the current codebase.

Humanizer

When writing with AI, I get annoyed by obvious AI tells: em dashes, repeated openings and the other patterns people spot from miles away. That is why I use Humanizer. It does a good job of removing them.

My advice on top of that skill is to add your own flavour. Ask AI to look at emails and instant messages you have sent and extract your writing style. Humanizer can remove generic patterns, but your own examples make the writing sound like you.

Example: use Humanizer with your own writing style

Goal

Rewrite a project update so it sounds like me.

Prompt

Rewrite this project update using the writing style in [YOUR WRITING SAMPLES]: "I am thrilled to announce that we have successfully completed this important milestone. Moving forward, we will continue to leverage our learnings and collaborate closely to drive meaningful impact." First compare it with the samples, then remove generic AI phrasing while keeping the facts, structure, and level of directness.

Outcome

Before

I am thrilled to announce that we have successfully completed this important milestone. Moving forward, we will continue to leverage our learnings and collaborate closely to drive meaningful impact.

After

We finished this milestone. Next, we will use what we learned and keep working together on the next steps.

It removes inflated language like "thrilled," "successfully," and "drive meaningful impact," making the update more direct and natural without changing its meaning.

Make them yours

Read them, check the audits, and change them for the work you actually do. Good instructions give an agent better direction, and your own changes make them useful for you. I hope this helps you find a few that work for you.