Drag the heart. Swipe to begin.

Module B

Putting itto work

The Kidzink heart character
Module BAI enablement workshopWith Adnan
From chat window to finished work

What we will cover

0 min
B0

How it actually works

15 minutes
Same box as Module A. One new source at a time

It is still just a boxA skill is just writingA tool is a scriptThree sources, one promptMemory is text too 

LLM
Its reply
0tokens in, this turn
⚙The script runs
It asks for itThe answer comes back, as text

Tap a source to switch it off.

From Module A. Text in, text out, one word at a time. Nothing we add today changes that.

A skill is a written procedure, added to what you send. It is not a program. It is more text.

It cannot run anything itself. It asks, something outside runs it, and the result returns as text.

Three places the text came from. One prompt. One reply.

Not a new mechanism. Saved from an earlier chat, pushed back in with everything else.

Which memory gets pulled in?

Everything gets a positionYour question lands too 

Every note and past chat becomes a position. Things about the same subject land near each other. That is a vector database.

Whatever sits nearest gets pulled in and pasted into the prompt. Drag the question and watch what changes.

Pasted into the prompt

Nothing yet. Ask a question.

Drag the red dot
Skill, tool, memory: all in. One answer out

The loop is the harness

MemorySkillsTools
LLM
Finished?
Not finished? It goes again.
The answer
Pass 0

The loop around the model is the harness.

B1

Files, projects and artifacts

15 minutes
Claude → Projects

A project, not a pile of chats

Claude projects page: chats and tasks in the sidebar, and the projects tabs
Files, projects, artifacts

A chat starts empty. A project accumulates.

01
Attach PDFs, spreadsheets, drawings and scans. It reads them directly.
02
A twenty-page quote is one attachment, not twenty copy-pastes.
03
A project is a workspace that accumulates. A chat starts empty every time.
04
A project per client or job. Four or five, not forty.
05
Three documents in every project: a one-page brief, examples of good work, the constraints.
Conversation 1
Starts empty
What it knows before you type0%
Files, projects, artifacts

Make the project remember

01
A correction you make twice belongs in that third document.
02
The same prompt inside and outside a project, run side by side.
03
A brief chat at the start of each project, referred back to when new work begins.
04
Name chats so you can find them in six months. Date, then topic.
05
On Team plans a project is shared. Same context, same knowledge, everyone.

Six months from now, which one was it? Tap each name to fix it.

Files, projects, artifacts

Artifacts are real files

01
Real files: spreadsheets with working formulas, decks, documents, web pages, code.
02
Change one section instead of regenerating the whole thing.
03
Outline, expand, polish. Faster than writing linearly, and you keep control.
04
Copying text out to reformat it? It should have been an artifact.
Al Noor proposal · artifact
Your edit to “Scope” is gone.
Claude → Artifacts

Every artifact, in one place

Claude artifacts page listing an artifact with a lock icon
B2

Research

5 minutes
Ask normally, or brief a research task

Search, or deep research

01
Ask normally and it searches, with citations you can open.
02
Deep research runs many searches, follows threads, returns a report.
03
Brief it like a research task: the segment, the question, what to weigh.
04
Vague in, unusable out. Name the constraints.
05
Every claim still gets opened and read. Module A applies here too.
B3

Skills

10 minutes
Written once, used by everyone

A skill is a written procedure

01
A procedure it follows every time. A folder with a SKILL.md in it.
02
Written once, used by everyone, same shape of output every time.
03
Fix it once and everybody gets the fix. It does not live in one person's head.
04
Built-in skills already handle spreadsheets, decks, documents and PDFs.
05
Ours first:
Quote comparisonSpec to BOMCase study formatSOP format
quote-comparison/SKILL.md
Compare supplier quotes.
1. One row per supplier
2. Columns: price, warranty
3. Flag anything missing
4. Cheapest is not the answer; say why
Skills

Plugins bundle skills, commands and connectors, so a way of working installs in one go.

One install. The whole way of working.

Skills

Commands

Connectors

procurement-plugin
0skills installed
0commands ready
0connectors on
Two ways in

Adding a skill, two ways

B4

Connectors

5 minutes
Reading and acting inside a system

What do we let it touch?

01
Reading and acting inside a system instead of copy-paste.
02
Fewer steps, fewer transcription errors, and one clear question: what do we let it touch?
03
Central provisioning on Team and Enterprise, not installed one by one.
04
Read-only first. Writing comes later, and with somebody watching.
05
What we connect, what we deliberately do not, and who signs off.

Set the access, system by system

0systems it can read
0it can write to, each needing sign-off
Two views

What's on, and what's out there

From chat to agent

One toggle, chat becomes agent

Claude composer with the Chat and Cowork switch, Project or folder, and Manual
B5

Agents, and the setup that makes them work

20 minutes
Agents, and the setup

Chat gives you text. An agent does the work.

01
Chat gives you text. An agent does the work.
02
Chat answers questions. An agent completes the task and hands you the files.

“Compare the three supplier quotes in /Quotes and draft the purchase order.”

Which one for which job

Chat, cowork or code?

💬

Chat

Questions, drafts, rewrites. Text back.

🗂

Cowork

Works through your files and hands back finished ones.

⌘

Code

Scripts and pipelines, when the answer must be computed.

Tick what your task does.
Skip it and everything comes back flat

The setup stops it being generic

01
One dedicated working folder, not the whole drive. Context, projects and output inside it.
02
Three files it reads every session: who we are, how we write, how we structure documents.
03
Global instructions: the standing brief you would give a new hire on day one.
04
A project per recurring job, so each run builds on the last.
05
Describe the output like briefing a person. No special syntax, no clever phrasing.

The working folder

📁 Kidzink-AI/
├─ 📁 context/
│  ├─
│  ├─
│  └─
├─ 📁 projects/  al-noor/ · weekly-supplier-report/
└─ 📁 output/
Tap a file to look inside.
Agents, and the setup

It works in parallel. Check its dates.

01
It works in parallel, not one file at a time.
02
It resolves vague deadlines into real dates. Check a few of those by hand.
03
A live run on one of our files, start to finish.

Six quotes to read

One at a time0s
In parallel0s

What it turned into dates · tap to resolve

Claude → Scheduled

A prompt, and a time

Claude scheduled tasks page with ready-made tasks and their schedules
B6

Unattended runs

20 minutes
A prompt and a time. No code

It reasons through each run

01
A prompt and a time. No code, no integrations, no third-party tools.
02
Ours to start with: a morning pass over mail and calendar, a weekly scan of what changed.
03
Wiring tools give the same output for the same input. This reasons through the steps each run.
04
So it can make a judgement call, and two runs are never identical.
05
On your machine, only while it is awake. Cloud runs continue with the laptop shut.

The 8:00 morning pass

A wiring tool

A scheduled agent

On this machine: runningIn the cloud: running
Unattended runs

What a scheduled run can and cannot do

01
A cloud run needs its own access. Know what it can and cannot reach.
02
Nothing carries between runs unless you write it somewhere that persists.
03
Secrets go in the environment panel, never in a file that stays on your machine.
04
Hourly is the floor. You cannot schedule every fifteen minutes.
05
Every run counts against usage. Keep the prompt tight.

How often?

15 min30 minHourlyDailyWeekdaysWeekly

Usage a month
Run 1Checked supplier emails. Two delays found.
No pause-and-ask mid-run

Approval sits between two runs

01
Always test it by hand before you schedule it.
02
The run log is your receipt. Every step, every decision, with timestamps.
03
Everything happens under your name. Commits, messages, replies.
04
No pause-and-ask mid-run. Approval sits between two runs, not inside one.
05
One drafts and posts it for review, a person approves, a second one executes.
Run 1drafts, posts for review
→
Youapprove or reject
→
Run 2executes
Unattended runs

Trigger a run from your phone and come back to finished files.

Scheduled

Weekly supplier reportEvery Monday, 7:00 · last run 7 days ago
B7

Visual work

10 minutes
The work has moved upstream

The design system comes first

01
Brand, colours, fonts, voice, references. A couple of hundred words, not three lines.
02
Set up once for us, then every generation is on brand from the first draft.
03
Short prompts once the system is right. The work has moved into the system.
04
Decks, one-pagers, carousels, tender covers and prototypes, without Figma.

“A one-pager for the Al Noor tender.”

Image
Al Noor Primary School

Shaded play for ages 5 to 11, built around the trees you already have.

Timber climber · rope net tower · sensory garden

Read the proposal
3 linesof instructions, so it guesses the rest

Iterate for free before you prompt again

Change the section, not the page

01
Built-in controls and direct editing first. Those cost nothing.
02
Never regenerate a whole page to change one section. Ask for the section.
03
Save what works as a template, so you never start from zero twice.
04
Export to PDF, PowerPoint or HTML, or hand it off to be built properly.
05
Arabic layouts keep native sign-off. None of this removes that.
Heading sizefree
Templates:
B8

Who goes further

5 minutes
Where the answer must be computed, not written

Ask for a script, not an answer

01
Coding agents: repositories, scripts, pipelines. Where the answer must be computed.
02
Ask for a script, not an answer, when the number has to be right and repeatable.
03
Two or three people per team go deep. Everyone else uses what they produce.
04
Staying with chat, files and skills is a legitimate outcome of this workshop.
05
The optional build track: what it covers, and who it is for.

“What is the total of every supplier quote this quarter?”

Ask for the answer

Ask for a script

total_quotes.py sum(q.total for q in read_quotes("Q3/"))

Two or three go deep.
Everyone else uses what they build.

Any
questions?

Thank you. Next, for some of you: the optional build track.

Fling the characters around while we talk.