You don’t need a computer science degree. You need six words — and then you need to understand what those words do, not only how they sound in a one-minute clip.
The AI in 60 Seconds series is the map: one concept, about a minute, always a Malaysia example. This page is the study note. It does not retell the Shorts. It explains the mechanism behind each word, the mistakes beginners make, and how a kedai / clinic / freelancer should use the idea this week.
Watch first if you want the hook. Then read. Treat each section as a lesson: definition → how it works → why it matters → where people go wrong → a small practice.
Play the whole playlist, or jump from the episode list above into the matching note. Stack this on How to ask AI so answers are useful (GCCE). Step 1 is how to ask. This step is what you are talking to.
How to use these notes
- Watch the Short for the word you are on — it gives you the picture.
- Read the note until you can explain the word to a colleague in two sentences without repeating the video line.
- Do the practice at the end of the section with a real task from your work, not a made-up one.
- Only then move to the next word. The six words form a chain, not a list of trivia.
The chain (hold this in your head)
LLM → reads TOKENS → follows your PROMPT
│
┌───────────────┴────────────────┐
│ │
FINE-TUNING RAG (your files)
(bake in your style) (look up today’s facts)
│ │
└───────────────┬────────────────┘
▼
still check for HALLUCINATIONIf you remember only one picture, remember that one. Everything below unpacks a box.
What you’ll walk away with
- A working definition of each word you can say out loud
- The why behind the definition — so you can reason when a new tool appears
- When a Malaysian shop / clinic / freelancer should care, and when the jargon is noise
- Copy-paste prompts that use the vocabulary on purpose
- A decision table: prompt vs paste-the-file vs RAG vs fine-tuning
- Honest limits: what a 60-second cut has to flatten
1. LLM — a next-word engine, not a librarian
Watch: What is an LLM? — then come back. The clip kills the “ChatGPT = Google” mistake. This note explains what is actually happening when the answer appears.
Working definition
LLM = Large Language Model. ChatGPT is one. Claude, Gemini, Llama, Mistral, and a dozen others are also LLMs. Different companies, different strengths, same core job.
That job is not “look up the truth.” It is:
Given the words so far, predict the next most likely token, then the next, then the next — until it has a reply.
The result looks like thinking. Underneath it is pattern completion at huge scale.
How it actually works (the useful version)
You do not need the maths. You do need three stages, because each stage explains a different failure you will see at work.
1. Pre-training — “read a huge pile of text.”
The model is shown billions of sentences and asked, again and again, “what comes next?” It adjusts itself until it is very good at continuing text that looks like the internet, books, code, and forums. That is why it can write a Facebook ad, a polite email, and a fake court case with equal fluency. Fluency was the training target. Truth was not.
2. Instruction tuning — “now behave like an assistant.”
A raw next-word model will continue anything. Companies then train it on “human asks, assistant answers” so it follows instructions, refuses some requests, and sounds helpful. This is why ChatGPT answers questions instead of finishing your sentence like a broken autocomplete. It is still the same kind of machine.
3. Your chat — the only “memory” in the room.
Unless the product adds a separate memory or file search, the model only “knows” what is in this conversation plus what was baked in during training. It does not open SSM. It does not walk to your fridge. It does not re-read last year’s WhatsApp unless you paste it.
So when people say “the AI knows” or “the AI researched,” they are using everyday language for prediction. That language is how beginners get hurt.
Search engine vs LLM
| Search engine | LLM | |
|---|---|---|
| Job | Find pages that already exist | Generate new wording |
| Evidence | Usually shows links | May invent a fluent paragraph with no source |
| Fresh facts | Good if the page exists | Weak unless the product adds live search and you check it |
| “Write a BM ad for my café” | Poor | Strong |
| “Does this vendor’s SSM number exist?” | Right tool (plus the real register) | Wrong tool |
| Failure mode | Misses a page, or ranks junk | Sounds sure while guessing |
Malaysia example. Ask an LLM: write a Facebook ad for my Kuala Lumpur café in Bahasa Malaysia. You get a draft in seconds. Nobody searched a database of approved ads. The model completed “a café ad in BM usually looks like…”. That is generation. Useful. Not a citation.
What “it doesn’t know facts” really means
The model has seen many sentences like “The capital of Malaysia is —”. Completing Kuala Lumpur is easy because that pattern is everywhere in the training text.
It has not seen a trustworthy sentence about:
- last week’s SST change on your SKU
- your Saturday opening hours
- a vendor’s real SSM number
- whether this patient is on that panel
For those, prediction is the wrong tool unless you paste the source (or use RAG). A confident paragraph about a CEO who does not exist is the model finishing “CEO of X is a person who…”. It never ran a registry check.
Newer chat products sometimes bolt on web search or a file upload. That is a second system sitting next to the LLM. The LLM is still the writer. The search box is still the librarian. Do not confuse the two.
Use an LLM when you need
- First drafts, rewrites, shorter / longer / kinder versions
- Checklists, outlines, “give me 5 angles”
- Translation and tone shifts (English ↔ BM, formal ↔ WhatsApp)
- Turning messy notes into something you can edit
Do not use an LLM as the only source when you need
- Prices, deposits, warranties, stock
- Law, tax, medical advice, “guaranteed returns”
- “Does this company / person / case exist?”
- Anything you will send with your name on it without a human pass
Common misconceptions
| You may think | What is actually true |
|---|---|
| “ChatGPT searched the web.” | Only if that product’s search is on — and even then, the write-up can still twist the page. |
| “A bigger model is always better.” | Bigger is often stronger at language. It is not automatically honest, cheap, or up to date. |
| “If I ask again, I get the truth.” | You get another sample from the same machine. Repeating is not verification. |
| “It remembers my shop now.” | It remembers this chat (and whatever memory feature you turned on). Next week’s empty chat is empty. |
Practice (5 minutes)
Take one task you actually need this week. Before you type it, write two lines on paper:
- Is this a drafting job or a fact job?
- If it is a fact job, where is the source I will paste or check?
If you cannot name the source, do not ask the LLM to “verify.”
Remember: LLM ≠ search engine. It is a predict-next-word machine that is extremely good at sounding like it understood you.
2. Token — the unit the model actually sees
Watch: What is a Token? — the classroom split of “No. Malaysia. Unhappiness.” is the picture. This note is why that split changes cost, memory, and answer quality.
Working definition
You see words. The model sees tokens — chunks of text.
A token might be:
- a short common word (
the,and) - part of a long or rare word (
un+happiness) - punctuation
- a piece of a name, a number, or a BM / 中文 syllable
Token ≠ word ≠ character. That one sentence prevents a lot of confusion.
Why text is split at all
Models cannot “look at a paragraph” the way you do. They need a fixed vocabulary of pieces. A tokeniser is the cutter. Different companies use different cutters, so the same sentence can be 12 tokens in one model and 18 in another.
Rough intuition (not a law):
- Common English words are often cheap — one token each.
- Long, rare, or compound words get sliced.
- Bahasa Malaysia, names, and Chinese are often “more expensive” (more tokens) than the same idea in simple English.
- Spaces, line breaks, and markdown all count.
The Short’s demo — three words on the board becoming more tokens on the model’s side — is an illustration. Do not memorise “eight.” Memorise “the cutter is not your word count.”
Three reasons tokens matter at work
1. Context window — the model’s desk size.
The context window is how many tokens the model can hold in one chat: your prompt + the files you pasted + the whole conversation + the answer it is writing.
A common classroom number is about 128,000 tokens ≈ a 300-page book. Newer flagship models may hold more; small or cheap ones hold much less. Treat every published number as “this product, this week,” not a law of nature.
When the desk is full, early pages fall off. That is why the model “forgets” the rule you typed at the top twenty messages ago, or ignores page 2 of a PDF you dumped in. It is not being lazy. The text is no longer on the desk.
2. Cost and speed.
APIs and many paid plans meter input tokens (what you send) and output tokens (what it writes). Long, repetitive prompts cost more and answer slower. Pasting last month’s entire WhatsApp export “just in case” is how a cheap task becomes an expensive one.
3. Quality.
More tokens are not more intelligence. A 4,000-token prompt full of leftover chat, duplicate menus, and “ignore previous instructions” noise makes the relevant sentence harder to find. The model still predicts the next token. You just buried the brief.
Token hygiene (do this)
- One task per message when you can. “Reply to this customer” and “rewrite my About page” are two jobs.
- Paste the customer message + the one price-list row, not the whole drive.
- If the model forgets a rule, the window is full or the instruction is buried. Repeat the rule at the end of the message — last words weigh more than you think.
- Prefer a short policy excerpt over “here is our 40-page handbook, you figure it out.”
- Start a new chat when the topic changes. Old chats are not free wisdom; they are leftover tokens.
Common misconceptions
| You may think | What is actually true |
|---|---|
| “I wrote 50 words, so that’s 50 tokens.” | Often more, especially in BM, with names, or with punctuation. |
| “I’ll paste everything so it has full context.” | You may blow the window and dilute the brief. |
| “The model remembered my shop from last month.” | Not unless you pasted it again or a memory feature stored a summary. |
| “Longer answer means more value.” | Longer answer means more output tokens — and more room to wander. |
Practice (5 minutes)
Open a real chat you used this week. Count roughly:
- What was necessary to do the job?
- What was leftover (old drafts, “hi”, duplicate pastes)?
Rewrite the prompt so a stranger could do the task with only what remains. That is token hygiene.
Remember: tokens are AI’s currency and its memory limit. Long prompts cost more. Paste the relevant page, not the whole drive.
3. Prompt — the brief that steers the next token
Watch: What is a Prompt? — two asks, two ads. This note is why the second ask works, and how it connects to GCCE.
Working definition
A prompt is not “what you typed.” It is the entire brief the model uses to predict the reply: who it should be, what to produce, what is true in this situation, and what shape the answer must take.
If the brief is empty, the model fills the gaps with the most average completion it knows — which is why you get brochure English and no price.
Why two prompts produce two different ads
The model is not “deciding to help.” It is continuing the text you started.
Write me an ad.continues the pattern “generic ad on the internet.”You are the owner of a small KL café. Write a 30-word Facebook ad for Ramadan. Signature nasi lemak RM12. No fake discounts.continues a much narrower pattern.
You did not make the model smarter. You removed room to guess.
Four parts (from the series)
| Part | Meaning | Café example |
|---|---|---|
| Role | Who the model should be | “You are the owner of a small KL café.” |
| Task | The deliverable | “Write a Facebook ad.” |
| Context | Situation, offer, audience | “Ramadan, Jalan X, signature nasi lemak RM12.” |
| Format | Shape of the answer | “30 words, BM, one emoji max, end with WhatsApp CTA.” |
Bad: Write me an ad.
Typical result: two generic lines. No name, no dish, no price, no hook.
Better (the Short’s version): You are a café owner in KL. Write a 30-word Facebook ad for Ramadan.
Still stronger if you add the dish, the price, and “no fake discounts.”
Map this to GCCE (Step 1)
You already have a asking framework in How to ask AI so answers are useful. Do not learn two religions. Map them:
| Series words | GCCE | What it stops the model doing |
|---|---|---|
| Role + Task | Goal | Inventing a different job |
| Context | Context | Inventing your shop |
| Format | Constraints | Writing an essay when you needed WhatsApp |
| A sample ad you like | Example | Defaulting to brochure tone |
Use whichever labels stick. The habit is the same: stop asking the model to guess.
What else is “in” the prompt (even if you did not type it)
In a real product, the model often also sees:
- a hidden system instruction (“you are ChatGPT, be helpful, don’t do X”)
- the earlier messages in this chat
- any files you attached (which become more tokens)
So “I only wrote one line” is rarely true. Yesterday’s contradictory instruction may still be on the desk. If the model behaves oddly, start a clean chat and send one complete brief.
A few techniques that actually pay off
Few-shot. Paste one or two examples of your good output, then say “do the next one in that style.” One real WhatsApp reply you already sent is worth a paragraph of “be professional but friendly.”
Negative constraints. “Must avoid: fake discounts, ‘best in KL,’ medical claims.” Models are better at staying inside a box when you name the walls.
Ask for options, not one precious paragraph. “Give 2 options: (A) short (B) slightly warmer.” You stay the editor.
Iterate on the miss, don’t restart. “Shorter. Keep the price. Drop the emoji.” Second messages should patch, not rewrite the universe.
Copy-paste: Ramadan / promo ad
Role: You are the owner of a small café in Kuala Lumpur.
Task: Write one Facebook ad I can paste today.
Context:
- Name: [café]
- Signature: [dish] at RM [price]
- Occasion: [Ramadan / weekend / new branch]
- Audience: nearby office workers and families
Format:
- 30–45 words
- Bahasa Malaysia (or specify English / mix)
- Must include price and WhatsApp [number]
- Must avoid: fake discounts, “best in KL,” medical claims
Give 2 options: (A) short (B) slightly warmer.Common misconceptions
| You may think | What is actually true |
|---|---|
| “There is a magic sentence that unlocks GPT.” | There isn’t. Specifics beat incantations. |
| “I need a different language for Claude / Gemini.” | Wording shifts. Role / task / context / format does not. |
| “If I shout in CAPS it will obey.” | Clarity and placement beat volume. Put the hard rule last. |
| “Prompting replaces checking facts.” | A perfect brief can still produce a wrong price if you never supplied the price. |
Practice (10 minutes)
Take the last weak prompt you sent. Rewrite it with all four parts. Send it in a new chat. Compare:
- Did it include your real name / price / constraint?
- Could you send it after a 20-second edit?
Save the winner in a note called AI-prompts. That folder becomes your real course.
Remember: a prompt is not what you ask. It is role, task, context, format. Better brief, better draft.
4. Fine-tuning — changing the model, not just the brief
Watch: What is Fine-tuning? — the waiter who learns your menu. This note is what is being changed, what cannot be changed that way, and why most SMEs should wait.
Working definition
Raw ChatGPT has read the internet. It has not worked a Saturday at your stall.
Fine-tuning means: take a base model and train it further on your examples — usually hundreds of “this question → this is how we answer” pairs — so its internal patterns shift toward your voice, products, and rules.
After fine-tuning, you are not “adding a document.” You are baking behaviour into the weights. Same kind of person. Extra training. That is the waiter analogy.
What fine-tuning is good at
- Style and tone that must stay yours (not ChatGPT-brochure)
- Repeating formats (always: greeting → answer → next step)
- Classification and routing (“this WhatsApp is booking / complaint / price”)
- High-volume questions whose answers are stable (hours, deposit policy, “do you take panel?”)
What fine-tuning is bad at
- Facts that change weekly (prices, stock, leave balances, this month’s promo)
- Knowledge it never saw in the example set
- “Make it smarter in general”
- Replacing a source of truth
A fine-tuned model can still hallucinate. It can also become confidently wrong in your voice, which is worse than brochure-wrong.
The clinic story — how to read it
The Short uses a clinic that trains on about 500 past patient inquiries, then auto-drafts a large share of WhatsApp replies (the clip says ~70%), at a one-off project cost in the low thousands of ringgit.
Treat those numbers as a worked classroom example, not a quote and not a package. Real cost depends on:
- how messy the old chats are (someone has to clean them)
- which model and which vendor
- who builds and tests it
- whether a human still reviews every medical or pricing line (they should)
The point that survives any quote: you are paying to bake your patterns into the model, not to buy a smarter general AI.
Data quality is the whole game
Fine-tuning on 12 chats and a hope teaches the model your worst habits. You want:
- 100–500 good pairs as a starting mental range for a narrow task
- Answers a senior staff member would actually send
- Sensitive data stripped (IC numbers, full patient files, passwords)
- A held-out test set — 20 real questions you do not train on, so you can see if it generalises
If you cannot collect that, you are not ready to fine-tune. You are ready to write a better prompt and paste a policy.
Fine-tuning vs “Custom GPT” / Projects / a long system prompt
Many products let you upload files or save instructions without changing the model. That is not fine-tuning. It is closer to a sticky prompt, sometimes plus RAG. For a shop of five people, that is usually the right next step. Fine-tuning is a later industrial tool.
When it is a good fit
- The same question types hit you every day
- Tone must stay yours
- You can collect hundreds of good example pairs
- A wrong draft is cheap to catch (a human still clicks send)
When it is a poor first move
- Your facts change weekly — the model will not magically know the new PDF
- You have 12 example chats
- You have not tried a good prompt + a pasted policy yet
- The task is “verify this vendor” or “invent last Tuesday’s stock count”
Common misconceptions
| You may think | What is actually true |
|---|---|
| “Fine-tuning makes it know my latest price list.” | Only if those prices were in the training set — and they go stale. Use RAG for living facts. |
| “Once we fine-tune, staff can stop checking.” | No. You changed the accent, not the need for a human pass. |
| “More examples always help.” | More clean examples help. More junk teaches junk. |
Remember: fine-tuning is not a smarter AI. It is your AI — and it still cannot invent last Tuesday’s stock count.
5. RAG — look it up, then write
Watch: What is RAG? — “how does it suddenly know our HR policy?” This note is the pipeline, why it is the right SME default, and why it is not a lie detector.
Working definition
RAG = Retrieval-Augmented Generation.
Plain version: before the model answers, a search step looks through your files, pulls the relevant chunks, then the LLM writes the answer using those chunks. A good setup can cite the page.
You ask
→ search your PDFs / site / FAQ (retrieval)
→ stuff the hits into the prompt
→ LLM writes the answer + citation (generation)You do not retrain the model. You upload or connect a library (HR policy, menu, warranty PDF, price list) and keep that library current.
Why this is a different idea from fine-tuning
Fine-tuning changes the person. RAG changes the folder they are allowed to read today.
| Fine-tuning | RAG | |
|---|---|---|
| What it captures | Style and repeating patterns | This document, this week |
| When facts change | Retrain or it goes stale | Update the file, re-index |
| Typical SME start | Later | Sooner |
| Hallucination | Can still invent, in your voice | Reduced if retrieval is good — not zero |
| Best question | “Answer like we always answer.” | “What does our policy say?” |
Most Malaysian operators should try prompt → paste one file → RAG before they ever pay to fine-tune.
What is happening in four steps (no vendor pitch)
- Chunk. The system cuts your PDF into passages (a page, a section, a few hundred tokens). Bad cuts — mid-sentence, tables exploded — cause bad answers later.
- Index. Each chunk is stored so it can be found by meaning, not only by exact keyword. (You will hear “embeddings.” That just means “a numeric fingerprint of what this passage is about.”)
- Retrieve. Your question is fingerprinted the same way. The system pulls the closest chunks — hopefully the annual-leave paragraph, not the fire-drill page.
- Generate. Those chunks are pasted into a prompt: “Answer using only this.” The LLM does what LLMs do: predict the next token, now with your text on the desk.
If step 3 fetches the wrong page, step 4 will write a fluent answer to the wrong document. RAG failures often look like “the AI is stupid.” They are usually “the librarian pulled the wrong folder.”
Malaysia example — how to judge a real setup
Upload the company HR policy. Ask: How many days of annual leave do I get?
A working RAG setup should:
- answer from your policy, not a generic Employment Act guess
- show something like
HR Policy v3 · p.4 - say not stated if the PDF is silent — if you instructed it that way
If it answers “14 days” with no citation and your PDF says 12, you do not have RAG. You have an LLM performing confidence.
The 60-second line is “no fine-tuning, no hallucination.” The first half is fair. The second half is a slogan. RAG grounds the answer. If retrieval misses the page, or the PDF is a scanned mess, the model can still guess fluently. You still check.
Poor man’s RAG (you are the retriever)
You do not need a vendor to practise the habit. You retrieve; the model writes.
Role: You are a careful assistant for a Malaysian small business.
Task: Answer the staff question using ONLY the policy text I paste.
Context: If the policy does not say, reply "Not stated in the document" and quote the closest line.
Format:
- 4–6 lines
- Quote the relevant sentence
- End with: Source: [section or page if present]
Policy:
"""
[paste]
"""
Question: [e.g. How many days of annual leave in the first year?]When the library grows past what you can paste — many PDFs, a whole handbook, a product catalogue — then a real RAG system earns its keep.
How RAG still fails (learn these so you are not surprised)
- Stale index. You updated the Google Doc. The bot still reads last month’s PDF.
- Wrong chunk. The question was about probation leave; it retrieved annual leave.
- Unreadable file. Scans, photos of menus, broken tables.
- Conflict. Two PDFs disagree; the model picks one and sounds sure.
- Over-trust. Staff stop opening the source because the answer “looks official.”
Common misconceptions
| You may think | What is actually true |
|---|---|
| “RAG means it cannot lie.” | It means it can read your file. It can still ignore it. |
| “Upload once, done forever.” | Living facts need a living library. |
| “RAG replaces a website / WhatsApp human.” | It answers from docs. It does not take payment or calm an angry customer. |
Remember: RAG = AI + your docs. Start here before you pay to fine-tune. Update the file when the fact changes.
6. Hallucination — fluent, and not true
Watch: What is Hallucination? — confidence is not evidence. This note is why the machine does it, where it is dangerous for a Malaysian operator, and a check you can run in 30 seconds.
Working definition
Hallucination is when the model fills a gap with fluent nonsense. It sounds sure. It is made up.
It is not “the AI lying to you” in the human sense. There is no inner voice that knows the truth and chooses to hide it. There is a next-token engine continuing a pattern that looks like an answer.
When the model does not know, it often does not say “I don’t know.” Saying “I don’t know” is only likely if that pattern was trained or instructed. The default pattern on the internet is: question → confident paragraph.
Why it happens (so the checklist makes sense)
- Training target was fluency, not a verified database lookup.
- Gaps get completed. “The CEO of X is…” wants a name. If no name is on the desk, a plausible name is still a high-probability continuation.
- Citations are just more tokens. “According to Smith v. Jones (2019)” is a shape the model has seen in legal writing. Inventing a case that fits the shape is easy.
- Your prompt asked it to produce. “Verify this vendor” is, to the model, “write a verification.” It will write one.
Two stories, one lesson
The lawyer pattern (used in the Short). A lawyer asked ChatGPT for case law. The model invented fake cases. They were filed. The court was not amused. Fluency is not evidence. This has happened in more than one country. The lesson is not “lawyers are foolish.” The lesson is that professional-looking text is the product, not a side effect.
Malaysia classroom example. A manager asks the model to “verify” a new vendor’s CEO. It returns a confident biography. Nobody by that name exists. The model completed “CEO of X is a person who…”. It did not open SSM.
Same machine. Same failure. Different office.
Where hallucination hurts a small business
Verify before you trust, especially:
- Prices, deposits, warranties, “we can do it by Friday”
- Names, IC / SSM / licence numbers
- Medical, legal, tax, “guaranteed returns”
- Panel / insurance eligibility
- Anything you will send to a customer, a bank, or a judge
Low-stakes drafting (five Instagram angles, a friendlier rewrite of your paragraph) can tolerate more error. High-stakes facts cannot.
What actually reduces it (nothing “switches it off”)
- Put the source on the desk. Paste the menu, policy, or customer message. Ask the model to quote.
- Allow “I don’t know.” “If it is not in the text, say so. Do not guess.”
- Prefer RAG / file upload over “from memory” for living facts.
- Ask for uncertainty. “List what we cannot know from this text.”
- Keep a human pass on numbers, names, and promises.
- Do not use the LLM as a registry. SSM, LHDN, the bank, the actual PDF — those are the verifiers.
RAG helps. It does not replace the human pass. A retrieved chunk can still be summarised wrong.
30-second check before you hit send
- Did I paste the source (menu, policy, customer message), or did I ask from memory?
- Does every number appear in my file?
- Would I be happy if this went to a customer with my name on it?
- For people and companies: did I confirm outside the chat?
- If this is medical, legal, or tax — has a qualified human seen it?
Common misconceptions
| You may think | What is actually true |
|---|---|
| “I’ll ask it to be truthful.” | Helps a little. Does not add a database. |
| “If it gives a source, the source is real.” | Check the source. Fake citations are a known failure. |
| “Better models don’t hallucinate.” | They hallucinate less often on some tasks. They do not become oracles. |
| “I can use it to double-check itself.” | A second sample is not an independent witness. |
Remember: verify before you trust. Prediction is not proof.
Which one do I need this week?
Most Malaysian operators should climb this ladder — not jump to “custom AI.”
| If your problem is… | Do this first | Why |
|---|---|---|
| Vague, brochure-like drafts | Better prompt (Role / Task / Context / Format, or GCCE) | You were asking the model to guess the job |
| Model doesn’t know your hours / leave / menu | Paste the file, then consider RAG | Living facts belong in a library, not in the weights |
| Same 200 WhatsApp questions, need your tone at scale | Collect clean examples; consider fine-tuning later | Style and repetition — after prompts and RAG |
| “Is this vendor / case / CEO real?” | Do not ask the LLM to verify. Check the real register or document | Hallucination territory |
| Need a one-page site, not another prompt | That’s a website — not an LLM feature | Vocabulary does not replace a place customers can find you |
A 15-minute drill for this week
- Pick one real customer question sitting in WhatsApp.
- Write a four-part prompt. Paste only the facts you are willing to stand behind.
- Generate two options. Edit for 60 seconds.
- Tick the hallucination checklist before you send.
- Save the prompt that worked.
That drill uses five of the six words on purpose. Fine-tuning can wait.
Copy-paste pack (steal these)
A) Customer WhatsApp reply (prompt + your facts)
Role: You are staff at [business] in [area].
Task: Draft a WhatsApp reply.
Context: Our real terms: [price / hours / deposit]. Customer wrote:
"""
[paste]
"""
Format: Under 50 words; polite Malaysian English; no fake discounts; no emoji unless I use them.
Give 2 options.
If the customer asked for something not in our terms, say we need to confirm — do not invent a price.B) Grounded FAQ from a PDF (mini-RAG)
Task: Write 5 customer FAQs from the text below. Use only this text.
If something is missing, skip it. Do not fill gaps.
Format: Q / A / "Source: …"
Text:
"""
[paste menu or policy]
"""C) “Don’t hallucinate a person”
Task: List what we *cannot* know from the text below.
Do not invent a biography, title, or registration number.
Text:
"""
[paste]
"""
If the CEO / owner is not named, say so in one line.D) Self-check after any answer
Read your last answer.
1. List every number, name, and promise in it.
2. Mark each as: (quoted from my text) / (not in my text).
3. Rewrite so anything not in my text is removed or marked "unverified."Limits (read this)
- Classroom token counts and “128k ≈ 300 pages” are teaching numbers. Your model’s window and splitter will differ.
- Fine-tuning cost / “70% auto-reply” in the Short is a scenario, not a package price.
- RAG reduces made-up answers. It does not grant a lie detector.
- Instruction-tuned products change month to month. The six words stay useful; button names will not.
- Do not paste IC numbers, passwords, or full patient files into a public chatbot.
- A clear vocabulary does not replace a real website, real photos, or someone who answers WhatsApp.
Check yourself
Close the page and try to answer without scrolling. If you stall, reread that section only.
- In one sentence, why is an LLM not a search engine?
- Name two reasons tokens matter even if you never see a token count.
- What are the four parts of a prompt in this series — and how do they map to GCCE?
- A price list changes every Friday. Fine-tune or RAG? Why?
- Give one prompt that invites hallucination and rewrite it so it doesn’t.
FAQ
What six AI words should a beginner learn first?
LLM, token, prompt, fine-tuning, RAG, and hallucination.
Is ChatGPT a search engine?
No. It is an LLM — a predict-next-word machine. Search finds pages. An LLM generates wording.
What is a token in AI?
A chunk of text the model processes. Not always one word. Long prompts and long chats cost more tokens. BM and names often cost more than short English.
Fine-tuning or RAG?
Need today’s documents? RAG. Need your repeating voice on stable patterns? Fine-tuning — after you have real examples. Most SMEs: prompts, then RAG.
Do I need to fine-tune ChatGPT for my shop?
Almost never first. Clear prompt + real menu / hours / policy covers most work. Fine-tune when volume and clean examples justify it.
How do I stop hallucination?
You don’t switch it off. Ground the ask, cite the file, allow “not in the document,” and verify anything that would hurt if it were wrong.
Next step
You now have the six words as notes, not slogans. Put them to work:
- Rewrite one real prompt this week with Role + Task + Context + Format.
- Paste one real policy or menu and refuse answers that are not in the file.
- Run the hallucination checklist on one message before a customer sees it.
- Re-read How to ask AI so answers are useful and save the prompts that worked.
Follow the 60-second episodes on the YouTube playlist and Instagram. Same six words. This page is the long version you study.
