Set up your workspace and course dataset

Module 1, Lesson 5 · 5 min read · Updated 21 September 2026

Module 1 · Lesson 5

Everything from here uses the same practice CVs and the same kind of tool. Fifteen minutes now, and the rest of the course is hands-on.

1. Download the practice CVs

You get 500 made-up CVs for one type of job, a job brief, and a shortlist of 20 chosen by a person. Nobody in the file is real. The email addresses use a reserved domain that cannot receive mail. Reuse all of it however you like.

The CVs are deliberately messy, because tidy ones teach you nothing:

What is in there	Why
455 normal CVs — 300 PDF, 155 Word	The baseline any tool should handle
40 two-column CVs with a sidebar	The most common reason a tool reads a CV wrongly
3 scanned CVs	The worst case. These are photos, not text
2 password-protected files	So your count will not match. That is the point
5 different date styles	Dates are the field tools get wrong most often
59 contractors, some with overlapping jobs	Often read as duplicate or impossible work history
52 career gaps, 58 career changers	The people most often lost by over-filtering
35 whose relevant work is only in a projects section	Rewards searching the full text, not just the fields
`manifest.csv` — the answer key	Use it after the exercises, not during

Download the practice CVs (ZIP)

2. Get a tool that can read them

You need something that can take a few hundred CVs, show you the data it pulled out, score them against rules you write, and explain a ranking.

Where this course describes a screen, it is describing HireGen on the free plan, which covers every exercise here. That is the one we use because we build it, and because it shows you the extracted data rather than only a score.

Open HireGen

You do not have to. Any tool works if it shows you the extracted data and lets you write your own rules. Several established hiring systems do. What will not work is a tool that only gives you a score, because half this course is about checking what sits underneath the score.

3. Load the CVs

Create a job and name it after the job brief in the folder.
Upload the whole /resumes folder. 500 files takes a few minutes.
Wait for it to finish before touching anything. Half-loaded results look like real results.
Check the count. You should see 498, not 500 — two files are password-protected and will fail. Find them in the log rather than assuming they were weak candidates.
Open any candidate and switch to the view showing the extracted data.

If something goes wrong

What you see	Why	What to do
Upload stops partway	Batch size limit	Upload 100 at a time
Fewer than 498 candidates	More files were rejected	Check the log; handle them separately
Missing names or sections	The scanned CVs	Expected. You study these in Module 4
Dates that make no sense	The tool misread the date style	Also expected. Note which layouts cause it
No way to see extracted data	The tool only gives scores	Switch tools before Module 4

4. Check your screen

You are set up correctly when you can see three things for one candidate: the original CV, the data pulled out of it, and a way to correct that data by hand.

One thing before you move on

Open three candidates at random — not the tidy ones — and compare the extracted data against the original CV. Do not fix anything yet. Just notice how often something is slightly off.

Noticing that is the skill this whole course is built on.

Common questions

Does AI screening actually save time?

It relocates time rather than removing it. Screening 400 applications manually at two minutes each is about thirteen hours, and in practice most teams spend four and read 150. AI screening takes that to roughly ninety minutes and covers all 400 — but it adds an hour of defining criteria you were not previously spending. The honest saving is large; the coverage gain is larger.

What is an applicant tracking system (ATS)?

An ATS is the system of record for hiring: it holds candidates, roles, pipeline stages and decision history. Everything else in an AI recruiting stack — parsing, matching, messaging, reporting — reads from it or writes to it. If two people asking the same question about a candidate get different answers, you do not have a working system of record.

Can AI make hiring decisions?

It should not, and in several jurisdictions decisions based solely on automated processing carry specific legal obligations. The workable line is that a system may order and annotate a pile of applications; a person decides who leaves it. Automate the reading, never the rejection.

Is AI in recruitment legal?

Using it is legal; using it without disclosure, records and human review increasingly is not. The EU AI Act classifies recruitment systems as high-risk, New York City requires an annual independent bias audit and candidate notice, and Illinois regulates AI video interviews. The rules differ by where your candidates are, not where you are.

Do I need a paid tool to follow this course?

No. Eight of the ten labs assume a screening tool, and each of those has a spreadsheet-only variant. The parsing layer is `pdftotext` and a spreadsheet; the scoring layer is a weighted sum. Doing it manually once teaches you more about where screening fails than any tool demo.

Check yourself

Four questions. Nothing is recorded anywhere but your own browser — this is for you, not for a score.

1A team deploys AI screening and their screen-in rate drops from 22% to 9% in a month. What is the most likely explanation?
A sharp drop in screen-in rate after deployment is the classic over-filtering signal. It looks like better filtering and is usually a criteria problem — and because rejections generate no feedback, nothing else will tell you.
2Which layer of the stack do vendor demos almost always skip, and why does it matter?
Demos start at matching and ranking because ranked lists look like intelligence. But a ranking is arithmetic over parsed fields, and a ranking built on corrupted input looks identical to a correct one.
3Why are false negatives more dangerous than false positives in AI screening?
A weak candidate reaching interview costs you thirty minutes and is visible. A good candidate rejected produces no event at all — so the natural feedback loop pushes you to tighten criteria in the one direction you cannot observe.
4A vendor cannot show you the parsed fields, only the match score. What does that tell you?
It is not necessarily disqualifying, but it tells you where to concentrate your testing. Half of this course is about inspecting what sits underneath the score.