What Are Copilot Skills — And How Do I Know They Actually Work?

If you’ve used Microsoft Copilot recently, you may have noticed it doesn’t just answer questions anymore — it does specialized work, the way a subject-matter expert would. That capability comes from skills: self-contained packages of instructions, procedures, and sometimes dedicated calculator scripts that Copilot loads on demand whenever a conversation matches what that skill is built for.

Think of a skill as a specialist Copilot can call in. A generic AI assistant might attempt to estimate a car loan payment in its head — and get it slightly wrong over a 60-month amortization schedule. A skill built specifically for that job instead runs the exact math through a dedicated calculator every time.

I didn’t want to take that on faith, so I picked one concrete example — a car loan and lease analysis skill — and used it as my test case for a bigger question: can Copilot Cowork actually refine one of its own skills, and can I trust the result? The loan/lease skill isn’t the point of the story; it’s just the example I reached for to prove the process out. And to be clear up front: I never touched a line of the skill’s code myself — Copilot Cowork did the whole nine yards, from copying the skill into a scratch workspace through editing, validating, scoring, and publishing it back, start to finish.

Example 1: Using the Skill As-Built

I started simple — handing the skill two real lease advertisements, exactly as a shopper would see them in an email or a dealer flyer:

For both, the skill correctly pulled apart the numbers that matter — total scheduled payments, total minimum cash cost, and a flag that the ad didn’t disclose enough (cap cost, residual, money factor) to calculate the true interest rate. That refusal-to-guess was itself a feature.

Metric	Compact Sedan	Luxury SUV
Monthly Payment	$329	$399
Term	36 mo	36 mo
Due at Signing	$0	$3,194.93
Total Scheduled Payments	$11,844	$14,364
Total Minimum Cash Cost	$12,514	$17,159.93
Cost as % of MSRP	51.07%	39.84%

Example 2: Finding the Gap

Then I pushed further and asked the harder question directly: “What’s the true interest cost, accounting for all the fees?”

The skill could get partway there — it estimated a residual value and money factor from published market lease data and computed the standard money-factor-derived rate. But when I asked specifically for the rate including the acquisition and disposition fees rolled in, Cowork had to walk through the math step by step each time rather than reading it straight off the calculator.

“That’s the kind of calculation that’s easy to get subtly wrong doing it manually turn after turn — exactly the failure mode skills are supposed to prevent, and exactly the moment Copilot Cowork decided the skill itself, not just the prompting, needed to change.”

---

Example 3: Copilot Cowork Iterating on Its Own Skill

Rather than keep redoing that math by hand every time it came up, Copilot Cowork stayed inside the same conversation and refined the skill’s calculator so it computes that figure natively, permanently — all of it done by Cowork itself, not by me editing any files:

1. Pulled the skill into a scratch workspace to edit safely, without touching the live version.
2. Added the new calculation — a “true interest rate, all fees included” field that rolls the acquisition fee, disposition fee, and rent charge into one annualized rate, reported alongside the existing money-factor-only rate.
3. Validated the change — structural checks and a static code check on the calculator script, both passing clean.
4. Scored the result — the updated skill came back at 93/100, graded “Excellent.”
5. Published it to the live skill folder, and confirmed the new field worked by re-running the same lease numbers — this time getting the true, all-fees-included rate directly from the calculator.

Beyond One Skill: One-Shot, Multi-Scenario Modeling

Skills aren’t limited to answering one small question at a time. Copilot Cowork can take a single, dense prompt — stacking several loan scenarios, a comparison table, a cash-vs-loan analysis, and even a bilingual amortization table — and produce every piece of it in one pass. Here’s the kind of consolidated prompt that works in a single shot:

Prompt Section	What It Asks For
Loan Scenario Analysis	Monthly payment, total interest, and total paid across 7 different principal/APR/term combinations — including a mid-term refinance
Payment Comparison Table	Monthly payment, total interest, total paid, difference vs. highest payment, and the cumulative return if that difference were invested at 8% annually
Loan vs. Loan	Side-by-side comparison of two specific loan offers
Cash vs. Loan Investment	Paying cash vs. financing, modeling the cumulative return of investing the monthly difference at 8%
Amortization & Return Table (Español)	A full amortization table — monthly payment, interest, remaining balance, and return on that balance at 10% annually

One prompt, five structured outputs, one of them in a second language — all computed precisely rather than approximated, and all from a single back-and-forth with Copilot Cowork.

---

How Cowork Fits Together With Skills

Concept	What It Means
Trigger matching	Skills carry a description with the phrases and scenarios that should activate them.
Dedicated tools	Many skills ship their own scripts, so math and lookups are computed precisely, not guessed.
Personal & reusable	Skills are available across future conversations once created.
Self-improving, safely	Skills are re-validated — scored, security-scanned, tested — before any change goes live, entirely by Cowork itself.

---

Skills are, in short, how Copilot goes from “knows a bit about everything” to “actually good at your specific thing” — and now that depth is arriving for personal Microsoft 365 users too.