AXHUB5 cards

Five words, and
the conversation opens up

When you read about AI, unfamiliar terms trip you up.
You don't need all of them.
Here are the five that come up most, in plain words.

For deeper definitions, see the AXHub glossary

Five words1 / 5

01

Prompt= just an "instruction"

It sounds like a grand technical term, but it's only the instruction you give the AI.
There's no special grammar.

The one knack —
write it the way you'd brief a new hire.
You don't need "prompt engineering" if you have that instinct.

Basic shape Two or three rules + one thing to avoid + one line on the tone you want.
e.g. "In three sentences, polite, and skip the excuses."

Note This deck is a glossary set — for practice, see lectures L1·L2

Five words2 / 5

02

Hallucination= a plausible fabrication

When the AI produces something untrue in a confident voice.
It's not a malfunction but a known trait, so there's no switch to turn it off.

So the answer isn't a technique but a habit.
Numbers, names, dates, links — check them against the source.
This one word alone unlocks half the news.

One line to remember Making things up is more common than saying "I don't know" — so the more it matters, the more you ask for the source.

Note A known trait of generative AI — the response is a usage suggestion

Five words3 / 5

03

Human in the loop= a person decides at the end

A design that inserts human review at the important points of an automated flow.
Call it semi-automation.

Not fully automatic, not fully manual — and this middle path performed better.
Teams where a person only checked the exceptions beat teams that approved every item, by more than double (+71% vs +30%).

Source Stanford Enterprise AI Playbook (51 deployments) — median productivity gain

Five words4 / 5

04

Agent= delegate, then check

If a chatbot is "ask and answer,"
an agent is "give it a goal and it runs several steps on its own" — delegate, then check.

The more convenient it gets, the more permissions matter.
Start by granting read-only; deleting, sending, and paying stay with a person to the end —
that one rule avoids most accidents.

Go deeper The 5 checks before you delegate are in card No.12 and lecture L5, with a real template.

Note Definition follows the glossary entry "agentic AI"

Five words5 / 5

05

Context window= the AI's working memory

There's a limit to how much the AI can hold and read at once.
As a conversation grows long, it forgets the early parts, or quietly summarizes them away.

That's usually why an AI gets strange at the end of a long chat.
For a task with rules you set, it's safer to start a new conversation and give the rules again.

Symptom and fix "Why did it forget the rule I gave earlier?" —
At the end of a long chat, that's normal. New conversation + paste the rules again.

Note A Meta executive's deleted-inbox incident was tangled up with this trait too (case library)

AXHUBclosing

Five words are enough,
the glossary knows the rest

Give an instruction (1), know it can make things up (2),
let a person see the end (3),
watch permissions when you delegate (4), start fresh when it drags on (5).

With these five, most AI writing reads without snags.

Deeper definitions — AXHub glossary

Next: card No.16 (rumor fact-check) · lecture L0 (play first) — axhub.net

AXHub card No.17 — definitions align with the AXHub glossary. Only the figures (+71%/+30%) cite a verified source; the rest are plain explanations and usage suggestions.