Team Onboarding

Working with AI

A practical introduction for the 2DOT4 team, from the basics to working with AI, not just talking to it.

2DOT4 · The House of Lightn = 2.417 · Internal 2026

Chapter One

Why now

The shift

Last year, AI could talk.
This year, it can do.

The leap of 2026 isn't smarter chat. It's AI that takes actions and completes multi-step work: read a file, look something up, draft the reply, finish the checklist. Less a search box, more a capable new hire who can operate your software.

And it takes the busywork off your plate: it doesn't replace your craft. Setting a stone is still you.

Microsoft's 2026 Work Trend Index reports a 15× year-over-year rise in active AI "agents" across a 20,000-worker survey.

The honest reality

88%

of organisations use AI somewhere

~6%

get a real, measurable bottom-line impact

The difference isn't the tool. It's how you use it: on the right work, with a human checking the output. Today is about getting us into that 6%.

Source: McKinsey, State of AI 2025.

What you'll leave with

One framework

How to ask

A simple recipe for prompts that work, and how to let AI write the prompt for you.

One habit

Trust, then verify

Five reflexes that keep AI accurate and safe with our information.

One commitment

One task a day

Pick a single recurring job and hand it to AI, starting this week.

The path today: how it works → how to ask → chat vs co-work → connect & customise → which model → the wider toolkit → make it real.

Chapter Two

How it actually works

What an LLM is

A super-autocomplete that read most of the internet.

It writes by predicting the next small chunk of text, over and over. Done at enormous scale, that single trick is enough to draft emails, summarise contracts, and answer questions.

The catch: it predicts what's plausible, not what's true. That's why it can sound completely confident and still be wrong.

A brilliant-cut diamond dispersing a beam of light into fine rays

Two ideas explain most of its quirks

The working desk

Everything in the conversation sits on one desk (the "context window"). It's large but finite: when it fills up, the oldest things slide off and it forgets how the chat began. Re-state important instructions in long chats.

A fixed education

The model's "schooling" was frozen at a cut-off date. It won't know last week's gold price unless it searches the web live. If a number changes daily, give it the number or let it look it up.

When it's confidently wrong · "hallucination"

Think of a confident intern who never says "I don't know."

You can't eliminate it, but five reflexes reduce it to almost nothing:

  1. Give it the source: paste the document; don't make it guess.
  2. Ask for citations: "show me where that came from."
  3. Let it off the hook: "if you're not sure, say so."
  4. Break big asks into steps.
  5. Verify anything that carries money, legal, or customer risk.

Your data · the part everyone asks about

On our business plan, what we type is not used to train the model, by default.

Reassuring

Commercial / Team / Enterprise plans: inputs and outputs are not used for training by default. Personal Free/Pro accounts ask you to choose, with an opt-out.

Still, the rule of thumb

Until our AI policy is published, don't paste: trade secrets (unreleased designs, supplier pricing), regulated personal data, or anything you wouldn't email an outside vendor.

Source: Anthropic privacy & commercial terms. Confirm which plan we're on; the default differs by plan.

Chapter Three

How to ask

A recipe that always works

Treat it like a brilliant new hire: tell it what you'd tell them.

Role
Who it should act as. "You are our procurement analyst."
Context
The background, and the why. "We're negotiating Q3 supply; margins are tight."
Task
The ask, as a verb. "Draft a vendor-comparison email."
Format
Exactly what the answer should look like. "A table, then a 3-sentence recommendation."
Examples
One or two samples of "good," when tone matters.

The same request · weak vs. strong

Weak

"Write a price-increase email."

Strong

"You are our sales lead. Write a warm, ~150-word email to a wholesale buyer explaining a 6% increase on 18k gold settings from 1 Aug: cite rising gold costs, reassure on quality and lead times, and end with an offer to discuss."

Then keep going, it's a conversation: make it shorter · add a table · more formal.

Working with Claude

Refine the request, not just the result.

Stuck on how to phrase it? Let Claude do the wording; you stay in charge of the intent.

  • Rough idea → clear brief: "Turn this into a clear request with role, context, task and format: [your messy one-liner]."
  • Sharpen what you have: "Here's my request: [paste]. What's unclear or missing? Give me a tighter version."
  • Let it interview you: "Ask me five quick questions, then proceed."

Chapter Four

Chat vs co-work

Two ways to use Claude

Chat

Ask a brilliant colleague

You ask, it answers. Drafting, explaining, summarising, brainstorming. Where most people start.

Co-work

Hand over the task

You give it a job and it takes actions, using your files and tools, and brings back finished work.

Same Claude. Different mode. The shift from "AI that talks" to "AI that does."

The everyday power features & where it's headed

Use today

  • Projects: a workspace that already knows your context (instructions + uploaded files).
  • Artifacts: editable, shareable documents, tables, charts and mini-apps, built as you talk.

Where it's headed · not switched on yet

  • Cowork: a desktop agent that works through your files and apps.
  • Claude for Chrome: acts inside your browser (asks before risky steps).
  • Claude in Slack: bring an AI teammate into a channel.
Claude for PowerPoint, Excel & Word · live demo

Ask Claude to build a deck, analyse a spreadsheet, or draft a document, and it produces the real file. Paste in an email thread and it drafts the reply. We'll try one live.

Chapter Five

Connect & customise

What is MCP?

A universal adapter for AI.

MCP (Model Context Protocol) is the open standard behind "Connectors." Like USB-C: one standard plug instead of a custom cable for every app. Sign in once; Claude can then read, and where you allow, act inside that tool.

Claudeasks & acts
ConnectorMCP
Your toolsDrive · M365 · Shopify
A central point of light connected by fine threads to many surrounding points

Connectors · our stack

The tools we'd plug in first.

Turn one on with a click (no code), sign in with your own account, then just ask in plain language.

Available to connect

Microsoft 365 · monday.com · Shopify

For example: "summarise last quarter's deals," "turn these notes into monday.com tasks," "what sold best on Shopify last month?"

Skills · teach Claude our way, once

A short written guide Claude follows automatically.

Available now · built-in

Claude already has skills for PowerPoint, Excel, Word and PDF. Ask, and it makes the real file.

Ideas to build for 2DOT4

Invoice / PO → a table to paste into Odoo · a customer quote in our format · a weekly production report · a vendor email in our house voice.

Anyone can write one in plain language, no coding required, then everyone reuses it.

Skills · what goes inside one

A short note, written the way you'd brief a new hire.

Name
What it's called. "Client prospect finder."
Description
When Claude should reach for it. This line is the trigger, so be specific. Vague, and it never fires.
Steps
How we do the job, in plain words. The procedure, written down once.
Example
One real input and the exact output you want back, so the format is unmistakable.
Files (optional)
A template, price list or form to reuse. Attach only if the job needs it.

No code. If a new hire could follow it, so can Claude. We'll build one together in a moment.

Chapter Six

Which model, when

Match the model to the trip

ModelLike a…Reach for it when
HaikuScooterQuick questions, short summaries, simple repetitive jobs. Fast & cheap
SonnetCar~80% of daily work. The sensible default
OpusTruckThe hard stuff: deep reasoning, complex multi-step work
Fable 5Freight trainThe most capable model, for the rare hardest jobs. Premium tier; may not be enabled on our plan.

Your plan includes a range of models, and for most work the default is fine. Don't overthink it.

Four diamonds increasing in size from small to large

Chapter Seven

The wider toolkit

One main brain + a few specialists

"What am I doing?" decides the tool.

Claude

Our default: writing, long documents, analysis, multi-step work.

Copilot

When you're already inside Word / Excel / Outlook.

Gemini

When you're already inside a Google file.

Also handy, by task: Perplexity (sourced web answers) · NotebookLM (your own PDFs) · Wispr Flow (dictation, ~3× faster than typing).

One rule: use only IT-approved tools, and never paste confidential information into any of them.

Chapter Eight

Make it real

Make it real · pick one task

One task each, starting this week.

Production

Photo of a finished piece → a QC checklist. Or translate a setting instruction for the bench, by voice.

Sales & Marketing

A new collection → 5 social captions + a buyer follow-up, in our brand voice.

Finance

A messy export → a categorised summary with the outliers flagged.

Procurement

Three supplier quotes → a side-by-side with the risks called out.

HR

A few bullets → a clear, inclusive job description.

Ops & Admin

Meeting notes → owner, task and due-date in seconds.

Start with non-confidential tasks. One champion per team. We'll keep a shared prompt library, and regroup in a couple of weeks.

Where we start

One framework. One habit. One task this week.

We don't sell diamonds. We engineer light. Now we engineer with it.

P.S. This whole presentation was made with AI.

2DOT4 · Above ground. Beyond standard.Champions + a 2-week check-in

←→ navigate · F full screen · N notes · M menu

Speaker notes