Where we are: Kestrel, week 1
The workspace you built in b1 is loaded: three role scorecards, the levelling and band sheet, the house interview guide, the hiring policy, the glossary, the house rules. No CVs. Kestrel is 400 people going to 600, which is 120 roles this year, and the first five open on Monday. You are the HR partner. Nadia S. is the recruiter, and between you that is the entire people team on this wave. This morning Dara O., who runs support engineering, sent you the whole brief for the Senior Support Engineer role: "Same as the last one but better. Someone senior. Can you get it posted today?"
The role scorecard: outcome, capability, evidence 7 min live
A job description describes a job. A scorecard describes what success looks like and how you would recognise it in a stranger. The difference matters because in b4 these exact lines become the screening criteria, and a criterion you cannot evidence turns into a vibe with a number next to it.
LiveThe four blocks of a scorecard3 min▶
Four blocks, in this order, because each one constrains the next. Outcomes first, because until you know what this person has to achieve you cannot say what they need to be able to do.
| Block | What goes in it |
|---|---|
| Outcomes - 3 to 5 | What must be true by month 6 because this person was here. Written as results, not activities. |
| Must-have capabilities - 4 to 6 | What they have to be able to do to hit those outcomes, each written so a stranger could show you |
| Nice-to-haves - as many as you like | Genuinely optional, never scored as a gate, kept in a separate block so they cannot creep upward |
| Anti-requirements | The things people ask for that are proxies. Named out loud so nobody quietly reinstates them |
Four to six must-haves feels uncomfortably few the first time. That discomfort is the point: a fourteen-item must-have list is not a high bar, it is an undefined one, and it will be scored on whatever the reader can actually see.
The list that was really a wish. A support team wrote eleven must-haves for a senior hire. When someone asked which of the eleven the current top performer would have failed at the point of hire, the answer was five. The list had never been a bar. It was a description of a person after two years in the seat, projected backwards onto strangers.
LiveWriting a capability so it can be evidenced3 min▶
Here is the whole skill in one contrast. Read both and ask yourself what you would actually do with each on a Tuesday afternoon with 24 applications open.
| Undefined | Evidenceable |
|---|---|
| Strong communicator | Has explained a technical failure to a non-technical customer, in writing, under time pressure |
| Customer-obsessed | Has held an escalation open against internal pressure to close it, and can say why |
| Self-starter | Has built a runbook or process nobody asked for, that other people then used |
| Senior | Has been the final technical escalation point for a paying customer, unsupervised |
Notice what happens to the rating. With the left column you rate the person. With the right column you ask for an example and rate what they did. The right column also tells you what to ask in the interview, which is why b5 becomes almost mechanical once this document exists.
Undefined criteria do not stay empty. When a criterion has no definition, raters fill the gap with the most visible signal available, which on paper is the writing itself: fluency, polish, an unbroken timeline, a familiar employer. Those travel with background rather than with ability. This is why structured, evidenceable criteria are not only the fairer choice - they are the more accurate one. You are measuring the job instead of the CV's editing budget.
Self-studyNice-to-haves and anti-requirements, kept honest2 min read▶
- A nice-to-have that gets scored is a must-have. If it earns points, it gates. Either move it up honestly or take it out of the rating entirely.
- Anti-requirements are the ones you have to write down. For Kestrel's Senior Support Engineer: a computer science degree, a named list of prior employers, an unbroken employment history, "native English", availability to start within two weeks, and a specific ticketing tool that takes a week to learn.
- Each one gets a replacement, not a deletion. The degree becomes the capability it was standing in for. The ticketing tool becomes "has worked a queue with SLAs". The point is not to lower a bar, it is to state the bar in job terms.
- Show the anti-requirements to the hiring manager. Half of them came from that manager, and the conversation goes far better when the list is on a page rather than in your tone of voice.
What AI is genuinely good at here, and what it must not touch 5 min live
This session is the strongest case for AI in the whole track, and it is worth being precise about why. Nothing here is a decision about a person. It is structure work on a document, and structure work is exactly what the machine is good at.
LiveThree things it does better than a tired human at 6pm3 min▶
- Ramble to structure. A manager's voice note - eight minutes, three tangents, one good sentence about escalations - becomes a draft scorecard with outcomes and capabilities in the right blocks. You are not asking it to know the job. You are asking it to sort what the manager already said.
- Spotting the requirement that is really a proxy. Ask it to flag every line that measures background rather than ability and it will catch the degree requirement, the years-of-experience number, the "cultural fit", the coded adjectives. It is unembarrassed about this in a way colleagues sometimes are not.
- Generating the evidence question. For every capability, ask "how would a candidate show me this?" and you get the work sample, the interview question and the reference question in one pass. That is three later sessions of work pulled forward into tonight.
Kestrel, ten minutes in. Dara O.'s actual voice note contains one line worth gold: "the thing that killed us last year was nobody owning the enterprise escalations end to end." That is outcome one. It also contains "someone from a proper software company", which is anti-requirement one. Same recording, and a structured pass separates them in seconds. A tired human at 6pm writes both into the ad.
LiveThe two lines it must not cross2 min▶
Both failures are quiet, which is what makes them worth naming before you start.
- It must not invent a requirement the manager did not ask for. Ask for a Senior Support Engineer scorecard and a model will happily produce a beautiful, plausible, entirely fictional set of must-haves drawn from every job ad on the internet. It reads as authoritative because it is fluent. Anything not supplied gets marked TBC - ask the hiring manager, and TBC is a perfectly good thing to send back to Dara O.
- It must not decide the bar. How many years, what level, what is worth rejecting for - those are Kestrel's calls, made by Dara O. and by you, and they get recorded. The model drafts the sentence. It does not set the threshold, and it certainly does not set it silently by picking a number that looked normal.
The invented five years. A first-draft scorecard came back with "minimum 5 years in enterprise support" in the must-have block. Nobody had said five. The manager, reading it, assumed a colleague had decided it and left it in. By the time anyone asked, forty applications had been screened against a number that came from nowhere. Invented requirements do not stay drafts for long - they get inherited.
The job ad: what narrows a pool and what widens it 5 min live
The scorecard is internal. The ad is the only part a candidate sees, and it does more filtering than any screen you will ever run, because the people it puts off never enter the numbers. You cannot audit a pool for the applications you did not receive.
LiveThe six lines that narrow a pool3 min▶
- The long must-have list. Fourteen bullets does not signal a high bar, it signals an undefined one, and it deters people who read requirements as requirements rather than as an opening offer.
- Degree requirements that the job does not need. If the capability is what matters, ask for the capability. A degree line filters on who could afford one.
- Years-of-experience proxies. "8+ years" is a stand-in for a capability nobody wrote down. It also quietly filters on age, on career breaks, and on anyone who got good fast.
- Coded and in-group language. Rockstar, ninja, warzone, aggressive, "one of the lads", "work hard play hard". Each one is a signal about who belongs, and people read those signals accurately.
- No pay range. "Competitive, depending on experience" asks the candidate to guess and then negotiate from behind. It reliably costs you the people who cannot afford to gamble an interview process.
- Describing the last postholder. The single most common failure. The ad describes Marcus after three years in the seat, so it selects for people who already resemble Marcus, and the team gets a little more like itself every hire.
Same job, different sentence. A team replaced "8+ years in enterprise support" with "you have been the final escalation point for a paying customer" and the applications roughly doubled. Nobody lowered anything. The requirement had been there the whole time, hidden inside a number that only some people knew how to translate.
LiveWhat widens it, without lowering anything2 min▶
- Five must-haves, each one evidenceable. Straight out of the scorecard, in the same words you will screen with in b4.
- The pay band, published. It costs you nothing you were not going to pay anyway, and it buys you applicants who would otherwise have self-selected out.
- What the work actually is. A day, a week, the queue, the on-call, the two accounts. People apply to work they can picture.
- Plain adjectives about work, not about identity. Calm in an outage. Clear in writing. Comfortable saying "I do not know yet, here is what I am checking."
- An honest line about what this job is not. Kestrel's role is not a stepping stone into product engineering, and saying so up front saves everybody four weeks.
The 24. Kestrel posts the widened version on Monday. By week 4, 24 applicants have landed for Senior Support Engineer, nine of them from backgrounds the old ad had never reached. That is the slate you will meet in b4, and every one of those 24 is about to be screened against the criteria you write tonight.
Self-studyTargeting: who sees the ad is part of the selection2 min read▶
Most people think of an ad as passive: you post it, and whoever finds it applies. Paid and algorithmic distribution is not passive. When you target a campaign, or when a platform's delivery model optimises who to show it to, a decision about the audience has been made before anyone has the chance to apply.
- It sits inside the regulated scope. Targeted job advertising falls within the EU AI Act's high-risk employment category (Annex III point 4). Who an ad is shown to is treated as part of selection, not as marketing that happens to sit near a hiring process.
- Optimisation can narrow without instruction. A delivery system optimising for clicks or applications can learn to show a role to the audience that historically engaged with it, which is the same backwards-looking pattern that b4 is built around.
- So audit distribution the way you audit the screen. Ask where the ad ran, who it was shown to, and what the targeting parameters were. If nobody can tell you, that is the finding.
- Not legal advice. This is education for HR professionals; obligations vary by jurisdiction and by what your platforms actually do. Leader session a2 is the legal map, and your DPO owns the position.
Turn Dara O.'s three sentences into a scorecard ★ 14 min · everyone builds
This is the whole session in one artifact. You start with a shrug and finish with a document that b4, b5 and b6 all run on.
Open the Kestrel workspace from b1. The scorecards, bands, guides and glossary are the context; nothing personal goes anywhere near it.
Paste the vague request below exactly as Dara O. sent it. Do not clean it up first - the mess is the input you actually get.
Run the working prompt. Read the draft for one thing only: is any must-have there that Dara O. did not say? Delete it, or mark it TBC.
Take the TBC list back to Dara O. as questions, not as a document to approve. Five specific questions beat one "does this look right?".
Set the bar yourself, with Dara O., in writing, and record who set it. That number never comes from the model.
The proxy audit on your own live ad ★ 8 min · build your own
Kestrel is practice. This is the one that changes something this week. Open a job ad you are running right now, or the last one you posted.
Paste the ad in and mark every line that is a proxy: degree, years, employer names, "native" anything, unbroken timeline, coded adjectives, culture-fit language, availability windows, tool names that take a week to learn.
For each marked line write two things: what capability it was standing in for, and how a candidate would evidence that capability instead.
Rewrite three of them. Three, not all of them - three you can defend to the hiring manager tomorrow beats fifteen you never send.
Check the pay line. If it says "competitive", find out what stops you publishing the band, and write down the actual reason you are given.
Ask where the ad is being distributed and on what targeting. Note whether anyone can answer.
The line nobody could explain. An audit turned up "must be able to start within 4 weeks" in the must-have block of a role that had been open for five months. It filtered out everyone on a three-month notice period, which in that market meant everyone currently employed at a competitor. It had been copied from an urgent backfill two years earlier and had ridden along in the template ever since. Nobody defended it. Nobody had ever read it either.
Try it yourself - this week ◐ 30-45 min total
- Finish the Kestrel Senior Support Engineer scorecard: outcomes, four to six evidenceable must-haves, nice-to-haves in their own block, anti-requirements with replacements. Save it in the workspace - b4 uses it directly.
- Do the same for one of your own open roles, and count how many of your must-haves you can name an evidence source for. The count is the finding.
- Run the proxy audit on a live ad and rewrite three lines. Send them to the hiring manager as a suggestion, not as a correction.
- Ask one hiring manager the Marcus question: "if the person who just left applied today as a stranger, would this ad have found them?" Write down the answer.
- Find out where your ads are distributed and on what targeting parameters. Bring the answer, or the silence, to b3.
Official sources covered
Taught from public regulation, published selection research, and the house rules you wrote in b1. This course is education for HR professionals, not legal advice; obligations vary by jurisdiction and any actual legal position belongs to your own counsel and DPO. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Your scorecard draft says "strong communicator". What is the fix?
Rail 2. Undefined criteria do not stay empty - raters fill them with the most visible signal, which is writing polish, and polish travels with background rather than ability. Naming the observable behaviour also hands you the interview question for free.
2 · The draft scorecard comes back with "minimum 5 years in enterprise support". Dara O. never mentioned a number. What do you do?
Two failures in one line: an invented requirement, and a threshold set by a tool rather than a person. Years of experience is also a proxy for a capability nobody wrote down. TBC goes back to the manager, and whoever sets the number is recorded.
3 · Marketing offers to run the Kestrel ad as a targeted campaign to a narrowed audience. What is the HR position?
Targeted job advertising sits inside the EU AI Act's high-risk employment scope (Annex III point 4). You cannot audit a pool for applications you never received, so ask where the ad ran, who it was shown to, and on what parameters. Not legal advice - check your own position.