CTO of Advancetrack. Engineering, management consultancy, accounting and academia.

4 minute read

Executive Summary

This week I learned how to build a Claude Skill - and the results surprised me.

The problem: Vendor due diligence. Every new app we consider at Advancetrack needs security review - data storage, T&Cs, accreditations. Each one takes ~30 minutes.

What I built: A Skill that automates the research and populates my standard template spreadsheet.

How I built it (45 minutes, once):

  1. Uploaded my existing completed template (from evaluating Claude itself)
  2. Explained to Claude how I use the spreadsheet
  3. Claude asked clarifying questions - I answered and uploaded our asset classifications
  4. Claude generated the Skill
  5. I tested the skill on a new vendor and cross-checked the results.

The result:

  • Claude researches and completes the spreadsheet: 2 minutes
  • I review and verify: 3 minutes
  • Total: 5 minutes vs 30 minutes previously

The key insight: Claude already knows what to look for (security issues, compliance gaps). The Skill defines the process - what to check, what format to output, when to flag concerns.

Accounting parallel: Imagine a Skill that reviews a bank statement PDF, cross-references to the ledger, creates a working paper, and flags discrepancies for human review. Same principle.

What repetitive professional judgment tasks could this approach work for in your practice?

Article

I cut a 30-minute compliance task to 5 minutes this week using Claude Skills. Here’s how. For those of you who don’t know, Claude is one of the leading Large Language Model AIs and it has developed strongly in the software development space. I am under a lot of pressure to deliver new features to the Advancetrack platform and have been looking at how I can improve my productivity using AI.

So what is a ‘Skill’? It’s a set of reusable instructions that tell Claude exactly how to approach a specific type of task. Claude already has the underlying knowledge (security vulnerabilities, accounting standards, compliance requirements) - the Skill defines the process to follow and the output to produce. Think of it as a detailed brief you’d give a competent new hire: “Here’s what to check, here’s the format I need, here’s when to flag issues.” A simple Skill in accounting might be to review a pdf of a bank statement, cross reference the statement to the related bank account in the ledger, create a working paper entry and flag to the accountant if there is a discrepancy.

I was reading a blog article from a developer about how to do this so I thought I would give it a go. One of the things I have to do for Advancetrack is to conduct due diligence on any new software we might want to use. One of my colleagues had asked me to do that on an AI note taking app she was looking at so I thought I would use this as a test case. I have a template spreadsheet for this that looks at things like where the data is stored, terms and conditions and security accreditations – one due diligence exercise takes around 30 minutes. What I did was load a copy of this (the one I completed some months ago for Claude itself) into the main chatbot interface of Claude (I use a private account with the option to allow Claude to train on my data set to off). I explained to Claude how I use the spreadsheet and asked it what it needed to know to define a skill so that I could use it to do this work for me. Claude came back with a list of questions which I answered. This included uploading a list of our information assets, their security classifications and targets relating to business continuity. Claude then generated a Skill and added it to its library – this took about a minute. I then asked Claude to try out the skill on the notetaking app by giving it the website url. Claude spent around two minutes researching the app and came back with the completed spreadsheet. Like any software developer, I then checked what Claude had generated for me – it was exactly what I would have concluded.

So in summary, my due diligence task has dropped from around 30 minutes to around 5 minutes – 2 minutes for Claude and 3 minutes for me to review its findings. All of this took me about 45 minutes to deliver as a repeatable, AI driven process. I will review new outputs until I am confident that the skill is consistently good.

The more I read around Skills, the more I can see that I don’t know – they were only introduced in October 2025 so there are a lot of people going up the learning curve on this and I am only just starting. I will continue to write about them as I learn more.