G'day there... this is how I think AI skills should actually be built.

Not by opening a SKILL.md file and trying to document the perfect process before you’ve done the work.

Start by working through the process with AI.

Then validate it, improve it and only package it as a skill once you know it works.

A real example in SharePoint

I recently demonstrated this using Copilot in SharePoint across two Microsoft Lists, this was from a real world client coaching session.

• Supplier invoices
• Purchase orders

The task was straightforward:

Take the PO reference on each invoice, compare it with the PO number in the purchase order list, identify what matches and explain what doesn’t.

I described the task to Copilot in natural language and asked it to generate a simple reconciliation report.

The result:

100 invoices reviewed.​
​66 matched.​
​34 exceptions.

▶️ Watch the video on YouTube to see how I worked through the reconciliation, improved the report and turned the finished process into a reusable skill.

But I didn’t turn it into a skill immediately.

That part matters.

A generated result isn’t automatically a reliable process

Before making the process reusable, I checked the report against the source data.

• Had Copilot used the correct fields?
• Had it matched the right records?
• Were the 34 exceptions genuine?
• Did the reasons given for those exceptions make sense?

The report looked convincing, but appearance isn’t validation.

This is where human involvement (you) remains essential. Someone who understands the business process must confirm that the AI has interpreted the data correctly.

If you turn an untested process into a skill, you haven’t created efficiency.

You’ve created a faster way to repeat a mistake.

Next came iteration

Once we knew the reconciliation was correct, we continued working with Copilot.

We improved the report’s design.

We added filters so someone could quickly switch between matched invoices and exceptions.

We also added a slide-out panel that explained why a particular invoice had failed to match.

At the same time, my site-level SHAREPOINT.md instructions were working in the background. They told Copilot where generated reports must be stored, so I didn’t need to repeat those rules in every prompt.

The logic was correct.

The output was usable.

Preview text: Start with the work. Prove the process. Then package what works.

Don’t start with the AI skill. Start with the work.

This is how I think AI skills should actually be built.

Not by opening a SKILL.md file and trying to document the perfect process before we’ve done the work.

We start by working through the process with AI.

Then we validate it, improve it and only package it as a skill once we know it works.

A real example in SharePoint

We recently demonstrated this using Copilot in SharePoint across two Microsoft Lists:

• Supplier invoices
• Purchase orders

The task was straightforward...

Take the PO reference on each invoice, compare it with the PO number in the purchase order list, identify what matches and explain what doesn’t.

We described the task to Copilot in natural language and asked it to generate a simple reconciliation report.

The result:

100 invoices reviewed.​
​66 matched.​
​34 exceptions.

▶️ Watch the full demonstration on YouTube to see how we worked through the reconciliation, improved the report and turned the finished process into a reusable skill.

But we didn’t turn it into a skill immediately.

That part matters.

A generated result isn’t automatically a reliable process

Before making the process reusable, we checked the report against the source data.

• Had Copilot used the correct fields?
• Had it matched the right records?
• Were the 34 exceptions genuine?
• Did the reasons given for those exceptions make sense?

The report looked convincing, but appearance isn’t validation.

This is where human involvement remains essential. Someone who understands the business process must confirm that the AI has interpreted the data correctly.

If we turn an untested process into a skill, we haven’t created efficiency.

We’ve created a faster way to repeat a mistake.

Next came iteration

Once we knew the reconciliation was correct, we continued working with Copilot.

We improved the report’s design.

We added filters so someone could quickly switch between matched invoices and exceptions.

We also added a slide-out panel that explained why a particular invoice had failed to match.

At the same time, our site-level SHAREPOINT.md instructions were working in the background. They told Copilot where generated reports must be stored, so we didn’t need to repeat those rules in every prompt.

• The logic was correct.
• The output was usable.
• The process was working.
• Only then did we ask Copilot to turn it into a reusable skill.

The progression matters

The process looked like this:

Conversation → Working process → Validation → Iteration → Skill

Now another person doesn’t need to understand:

• The prompts used to create the report
• The relationship between the two lists
• The reconciliation logic
• The report design
• Where the output should be stored

They can invoke the skill and run the process.

That’s where this becomes genuinely valuable.

A skill isn’t just a saved prompt.

It is a packaged way of working.

It captures the instructions, decisions and lessons discovered while completing a real task—and makes that knowledge available to other people.

Don’t automate what we haven’t understood

There’s a temptation to begin every AI initiative by asking:

A better question is:

We start with one real process.

We work through it conversationally.

We check the result.

We identify where the AI needs more context and where human judgement is still required.

We improve the process until it is dependable.

Then we package the “how” into a skill that everyone else can use.

Want to see the complete process in action? Watch the step-by-step demonstration on YouTube.

Don’t start by trying to automate everything.

Start with the work. Make sure it works. Then turn it into a skill.

Daniel

Only then did I ask Copilot to turn it into a reusable skill.

The progression matters

The process looked like this:

Conversation → Working process → Validation → Iteration → Skill

Now another person doesn’t need to understand:

• The prompts used to create the report
• The relationship between the two lists
• The reconciliation logic
• The report design
• Where the output should be stored

They can invoke the skill and run the process.

That’s where this becomes genuinely valuable.

A skill isn’t just a saved prompt.

It is a packaged way of working.

It captures the instructions, decisions and lessons discovered while completing a real task—and makes that knowledge available to other people.

Don’t automate what you haven’t understood

There’s a temptation to begin every AI initiative by asking:

A better question is:

Start with one real process.

Work through it conversationally.

Check the result.

Identify where the AI needs more context and where human judgement is still required.

Improve the process until it is dependable.

Then package the “how” into a skill that everyone else can use.

Want to see the complete process in action? Watch the step-by-step demonstration on YouTube.

Don’t start by trying to automate everything.

Start with the work. Make sure it works. Then turn it into a skill.

Daniel