PRESENTS · SESSION 4 · PART 1 · START WITH THE PROBLEM
SAT JUN 13, 2026 · 12:00 PM EDT
Start with
the problem
Think first, then AI. Get the problem right before you reach for a tool.
Today: how to take any HR problem and turn it into a spec.
CO-INSTRUCTORS
S
Sandhiya Thiruvengadam
Leader, HRBP & AI · Apexon
Q
Q Hamirani
Chief People Officer, Highlevel · Adjunct Faculty, LBS
Session 4 · Part 1 · Slide 1/18
YOUR CLIMB SO FAR
Look how far you've come.
From a chat box to a working tool. Chat, then Cowork, then Code.
WEEK 1
a 2x2 on a Friday
WITH CLAUDE
WEEKS 2 TO 3
a skill that writes like you
CLAUDE + COWORK
LAST WEEK
a working tool
built in an afternoon
COWORK + CODE
↩ HOME
Session 4 · Part 1 · Slide 2/18
Everyone's already moving.
In different directions, each chasing their own definition of done.
PAUL
TRIED
a JD rewriter
DONE MEANS
every edit explained,
not just made
NANCY
TRIED
a manager FAQ
DONE MEANS
the right answer,
the first time
ARJEN
TRIED
a reviews agent
DONE MEANS
reviews grounded in
one leadership model
ZAINAB
TRIED
a first-draft skill
DONE MEANS
output she sends
with light edits
OLIVIA
TRIED
an onboarding nudge
DONE MEANS
managers self-serve
the onboarding plan
...and the rest of you are mid-build. Everyone's "done" looks a little different.
That definition of done is the thing we put on the page today.
↩ HOME
Session 4 · Part 1 · Slide 3/18
Today, everyone gets on the ramp.
Not a new tool. You climb it with your problems, using whatever AI you already have.
WHERE IT LEADS
your spec
ANY TOOL GETS YOU UP
Claude · ChatGPT · Microsoft Copilot · Gemini
Pick the one your company already pays for. The climb is the same.
You start somewhere. You won't get it right every time. Each push up makes it better.
↩ HOME
Session 4 · Part 1 · Slide 4/18
Three weeks, every door.
Today, you point all of it at one problem.
Claude
Gemini
ChatGPT
Copilot
skills · routines · schedule · chat · cowork · code · and some of you, a git
We heard you. So today: one spec, written before any tool.
↩ HOME
Session 4 · Part 1 · Slide 5/18
The questions you all asked.
Every one of these is normal. You are not behind.
SO MANY QUESTIONS
?
?
?
?
?
all of them, completely normal
"Do I even need the terminal?"
"Can you show the skill behind it?"
"Does my computer have to stay on for it to run?"
"Why Gemini, not Claude?"
"How is a Gem different from a chat?"
"Skill, or artifact?"
"When do I split one big skill?"
"Do things carry between chats?"
UNDER ALL OF THEM, ONE QUESTION
Where do I start? Today, you start with a spec.
↩ HOME
Session 4 · Part 1 · Slide 6/18
The ramp is the shape.
The same three steps up: where it starts, what kicks it off, where it lands.
STEP 1 · INPUT
where it starts
the sheet, files, data
STEP 2 · TRIGGER
what kicks it off
a schedule, a status, a click
STEP 3 · OUTPUT
where it lands
a nudge, a doc, an answer
SAME THREE STEPS, WHATEVER YOU CLIMB WITH
Claude · ChatGPT · Microsoft Copilot · Gemini
A map, not a manual. You give the direction, not every footstep. Confidence is not control.
↩ HOME
Session 4 · Part 1 · Slide 7/18
You already learned the pieces
A spec just puts them in one place. You met these in Session 2, building agents in Copilot and Gemini.
BLOCK 1
Instructions
IN YOUR SPEC
What you want done,
and the rules it
must follow.
"Here's the job,
and here are my rules."
BLOCK 2
Knowledge
IN YOUR SPEC
What the AI must
see: your policies,
files, and data.
"Read this before
you answer me."
BLOCK 3
Memory
IN YOUR SPEC
What done looks
like, held so it
doesn't drift.
"Remember what
we agreed good is."
BLOCK 4
Tools
IN YOUR SPEC
What it's allowed
to touch, and where
the output goes.
"You may use Slack.
Send it there."
+ THE SHAPE
Input · Trigger
· Output
INPUT what it starts
from
TRIGGER what kicks
it off
OUTPUT where it
lands
from Week 3
THE HARNESS
A spec is a map, not a manual. You set the direction and where to grab the reins. Confidence is not control.
BUILT IN, NOT BOLTED ON
"Never rank down someone on parental leave" is a line in your spec, not something you catch after it ships.
↩ HOME
Session 4 · Part 1 · Slide 8/18
Start with a spec,
not a prompt.
A spec gets your definition of done out of your head and onto the page.
THE PART YOU ALREADY KNOW
A spec is a policy
document for your AI.
You write policies every day. You already
know how to do this.
A PROMPT
"Write onboarding emails."
Vague. The rest is still in your head.
A SPEC
Who. What input. What good looks like.
What's out of bounds. On the page, before you build.
"Prompting is delegation, not commanding.
The AI was never the bottleneck. My thin instructions were."
PAUL · COHORT 2 · LAST WEEK'S REFLECTION
↩ HOME
Session 4 · Part 1 · Slide 9/18
The spec first. The tool after.
The wrong tool automates the wrong problem. The right tool amplifies the right process.
THE OLD WAY · TOOL FIRST
Pick the shiny tool.
Then bend the problem to fit it.
You automate the wrong thing, beautifully.
TODAY · PROBLEM FIRST
Write the spec.
Then ask: which tool fits THIS?
The tool disappears. The work gets done.
ONCE YOU HAVE THE SPEC · WHICH SURFACE FITS?
One-off, personal
just talk it through
→ Chat or Cowork
Repeatable, shareable
runs again, others use it
→ Code, or an agent
Needs your real data
reaches your systems
→ connectors · Part 2
↩ HOME
Session 4 · Part 1 · Slide 10/18
Watch what happens with no spec
A real one. This happened while building today's deck.
a real conversation · no spec given
YOU
github.com/AlmostTechHR/Global-People-Labs.git
CLAUDE
"What would you like me to do with this repo?"
YOU
"read it"
CLAUDE
"Read it for what? Which files? To what end?"
YOU
"...we are going to work on week 4"
CLAUDE
"Which week 4? There are two copies. Doing what?"
Three rounds. Nothing built yet. Time and tokens, gone.
WHY IT KEPT ASKING
The AI isn't slow.
It's missing your
definition of done.
That lives in your head.
Until you hand it over,
every turn is a guess.
THE LESSON
Be close on what you
want before you spend
a single token.
↩ HOME
Session 4 · Part 1 · Slide 11/18
First question: should this even be AI?
Arjen said it in the room: "Not everything needs AI. Sometimes a plain integration is better."
HIGH IMPACT
LOW IMPACT
LOW EFFORT
HIGH EFFORT
QUICK WINS
Do these first.
JD drafts · meeting notes · FAQ bot
STRATEGIC BETS
Plan for these.
pay equity · attrition signals
CONVENIENCE
Nice, not urgent.
tidy a doc · reformat a list
RECONSIDER
Maybe not AI at all.
a plain system-to-system integration
YOUR FRAMEWORKS
You already have the toolkit:
IMPACT
spotting where AI helps
SCALE
choosing a vendor or tool
CLEAR
governing the data
The 2x2 is just the first cut.
YOU ALREADY DREW THIS
in Week 1.
Before you write the spec,
place the task on the grid.
If it lands bottom-right,
the best AI move is
no AI.
↩ HOME
Session 4 · Part 1 · Slide 12/18
A spec, written out in full.
Olivia's onboarding nudge. Notice: plain English, no code. Just the decisions only she can make.
OLIVIA'S SPEC · MANAGER ONBOARDING NUDGE
SPEC nudge each manager to finish their new hire's onboarding plan
INPUT the new-hire list (HRIS or a sheet) + each hire's manager
TRIGGER anyone not done by Friday
OUTPUT a short, friendly nudge in Slack (or email)
KNOWLEDGE your onboarding template + the 60-day policy
DONE every new hire has a filled plan within 60 days
NOT ALLOWED never nudge someone on leave; never show another
hire's details
Hand this to any tool. It now knows what you want, and what you won't allow.
WHY IT WORKS
Every line answers a question
the AI would otherwise guess:
INPUT → where do I look?
TRIGGER → when do I act?
OUTPUT → where does it land?
DONE → when am I finished?
NOT ALLOWED → never do what?
No guesses left.
No wasted turns.
↩ HOME
Session 4 · Part 1 · Slide 13/18
Your turn. Sketch one spec. On paper.
Pick one real task. No laptop. Just the shape. 7 minutes.
A WORKED EXAMPLE · A PEOPLE LEADER · COHORT 2
Nudge managers to finish onboarding
SPEC every new hire has a filled 60-day plan
INPUT the new-hire list (from the HRIS or a sheet)
TRIGGER anyone not done by Friday
OUTPUT a short nudge in Slack or email
KNOWLEDGE your onboarding template + policy
DONE a filled plan for every hire inside 60 days
NOT ALLOWED don't nudge anyone on leave
Notice: no code. Just what you want, and what you won't allow.
NOW YOURS · FILL THE BLANKS
SPEC I want ______________________
INPUT it starts from ______________
TRIGGER it kicks off when __________
OUTPUT it lands in ________________
KNOWLEDGE it must read ____________
DONE good looks like _____________
NOT ALLOWED never ____________
Reflection: where in your week is the thing only
you know how to define?
INTERMEDIATE → ADVANCED
Next: from framing to building. MCP is just how your spec reaches real data.
↩ HOME
Session 4 · Part 1 · Slide 14/18
So what can you actually build?
Real HR use cases, from teams in the field. AI now touches 43% of HR work. Not one of these a website.
Resume screening scored to a rubric, far faster
Candidate sourcing scans the whole web, not just boards
Onboarding Q&A answers new hires from your docs
Offer letters + interview kits drafted from the scorecard
Survey & sentiment themes from open-ended feedback
Policy + people Q&A answers straight from the handbook
ONE SPEC TO START
All HR. Each began as a spec, not a website.
BEYOND HR · THE SAME MOVE
Finance flags invoice mismatches · Marketing scrapes competitor prices · Ops sorts messy files by content.
↩ HOME
Session 4 · Part 1 · Slide 15/18
And we just built one. Live.
Olivia wanted managers to finish onboarding plans. No website. Just a spec, pointed at real tools.
HOW IT RUNS
The spec
written first, before any tool
A Google Sheet
the input, where it starts
The Friday check
the trigger, what kicks it off
A Slack nudge
the output, where it lands
THE REAL THING
the real sheet, and the AI reasoning over it
WHAT WE COULD NOT (YET)
On a free Slack plan, asking Claude inside the channel needs a paid plan.
So for now the channel self-serves, and the AI runs from Claude itself.
Same spec you just learned. Let's go see it live.
↩ HOME
Session 4 · Part 1 · Slide 16/18
No data yet? Draw it.
A move for any build: no real data yet? Ask Claude to generate a realistic sample. Olivia's nudge needed a manager sheet she didn't have.
HER SPEC SAID
INPUT
a list of managers, who is
onboarding, and their status.
The real sheet?
Not ready yet.
That stops most people here.
It doesn't have to.
THE MOVE
Ask Claude to generate a sample.
onboarding-sample.csv · generated, not real
MANAGER
NEW HIRE
DAY
STATUS
Priya R. Sam Okafor Day 12 done
Marcus L. Dana Kim Day 30 pending
Elena V. Tom Reyes Day 5 done
Raj P. Mia Chen Day 45 pending
Nadia H. Leo Park Day 18 done
Omar S. Ava Singh Day 60 pending
Fake names, real shape. Three pending, three done, enough to build and demo the nudge today.
Build against this now. Swap in the real sheet when it exists.
The spec never changes.
↩ HOME
Session 4 · Part 1 · Slide 17/18
Governance isn't a separate step.
It lives in your spec.
THE LINE THAT SAYS "NOT ALLOWED"
You wrote a data-sensitivity grid last week.
Those rules become lines in the spec:
never rank down someone on parental leave
never show payroll to a manager-facing tool
never store an individual survey response
never touch health or leave records
Paul sorted 8 data types. Nancy sorted 9. That work is your guardrail list.
Write the limit in before you build, not after.
WHAT THE AI MAY AND MAY NOT TOUCH
PUBLIC
policy docs · go ahead
INTERNAL
JDs · resumes · with care
CONFIDENTIAL
psychometrics · aggregate only
RESTRICTED
payroll · health · leave · keep out
The spec names the level. The AI stays inside it.
↩ HOME
Session 4 · Part 1 · Slide 18/18
SESSION 4 · PART 1 · COHORT 2
Cohort 2 is in the room.Session 4 — Deploy & make the case. These slides walk through it live with Q & Sandhiya. Cohort members sign in to open the deck; everyone else: catch Cohort 3.
ALREADY IN COHORT 2?
Welcome back. The deck loads behind the sign-in.
SIGN IN WITH YOUR INVITE →
WATCHING FROM OUTSIDE?
Cohort 3 opens later this summer. We'll save you a seat.
CATCH COHORT 3 →