Skip to content
Get in touch
Menu
Field Guide · level 1 · Start here · How it works, optional

How does AI know what words mean?

Why vague questions get average answers, and how to get yours.

5 min · try it 10 min · as of Oct 2026
Bring this to your team · Team workshopsA workshop where your team turns vague asks into specific ones. Get in touch
01

Fine, but not yours

You asked for a welcome email. You got one for anybody.

02

A map of meaning

Inside the AI, picture a map.

  1. Every word has its own place.
  2. Your question lands among them.
  3. Each detail you add moves it.
03

1. Similar words, close together

Words used in similar ways sit close. Dog and puppy are neighbours. Invoice lives across town.

04

2. Vague lands in the middle

A vague question lands in the busy middle, among everyone else's. It gets the answer most people get.

05

3. Details move it

Add who it's for, what it's for, and the limit. Each moves it toward your corner.

06

Vague in, average out.

Before you ask, add who it's for, what it's for, and the limit.

My pick for everyday AI

Claude logoClaudeAlso good: ChatGPT logoChatGPT

One conversation for quick questions and whole tasks. It asks before it acts.

All my picks
Try it · 10 minutes

Ask it twice

Made-up first, then your own. Nothing private.

  1. In Claude, ChatGPT or Gemini, ask the vague one.2 min
  2. In a new chat, ask the specific one.2 min
  3. Which could you send as it is?1 min
  4. Then fix one of your own: who, what for, the limit.5 min

The two questions

Vague:
Write a welcome email for a new starter.

Specific:
Write a welcome email for Sam, a new nurse on our night team. It's for Sam's first shift on Monday. Under 100 words, warm, and say who meets Sam at the door.

Good looks likeThe second one needs no blanks filled in.

In your assistant · as of Oct 2026 · Claude, ChatGPT, Gemini

Say it once, for every chat

Save what never changes. Add the job's details each time. Menus move often.

  • About youyour role and team
  • Readerswho reads your work
  • Limitsthe ones you always want

Standing instructions are for facts that rarely change. Leave out passwords, client details and anything private.

Claude

Calls it Instructions for Claude

About you
SettingsthenInstructions for Claude

Click your initials first. Every chat uses them.

One job
ProjectthenSet project instructions

Only that project's chats use them.

Check
SettingsthenMemory

What it remembers sits under Topics. Edit or delete there.

On Team and Enterprise, an owner decides whether memory is available.

ChatGPT

Calls it Custom Instructions

About you
SettingsthenPersonalizationthenCustom Instructions

With Enable customization on. On every plan.

One job
•••thenProject settings

Project instructions win over your general ones, in that project.

Check
SettingsthenPersonalizationthenMemory

Some controls vary by plan, region and workspace.

Gemini

Calls it Instructions for Gemini

About you
Settings & helpthenPersonal IntelligencethenInstructions for Gemini

Click Add, type them, then Submit. Every chat uses them.

Check
Personal IntelligencethenMemory

In the same menu. Memory needs a personal account, a Google AI plan, the US, age over 18.

Instructions are open to all users. On a work or school account, your Workspace admin decides who uses Gemini and how long chats are kept.

Quick check · three questions

Did it stick?

  1. 1On the map, which two words would sit closest?

    The answer: C Neighbours are words used in similar ways. Closeness comes from how words are used, not how they look or sound.
  2. 2Two colleagues each ask for “a project update email”. Both get much the same bland email. Why?

    The answer: A Nothing said whose project, for whom, or how long. Their own details would pull the two answers apart.
  3. 3Which addition moves “Write a job ad” furthest toward your corner?

    The answer: B Who it's for and a limit belong to your situation. Good, professional and long fit everybody's ad.
0 of 3 answered
A picture, not the wiringSide note
The map and the pin are our picture. Inside, each word is a long list of numbers, and the words around a word change what it means there. The next lesson shows how.
Made-up placesSide note
Where dog, puppy and invoice sit on our map is made up. Google's own course on embeddings says words used in similar contexts end up closer together.
Check what it remembersSide note
Standing instructions go into every chat, and memory can shape answers too. Read both now and then, and fix what is old or wrong. The memory lesson shows where.
Sources, checked 18 Oct 2026Source

Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.

Free kitThe 2026 AI StarterBring this to your teamTeam workshops · Get in touchBack to level 1, Start here