Session Summary · WSS-04 · Free Member Sharing
What Actually Happens in WSS-04: A 90-Minute Self-Check on the 5 Levels of Claude Adoption
You do not have a five-part AI problem. You have a one-part problem — and ninety minutes is enough to find out which part.
WSS-04, Where Are You on the AI Ladder?, is a free online sharing in the Webinar Short Sharing Series, written for SME owners, directors and department managers who have used AI in a chat box and nothing more. It sits before the hands-on programmes and answers one question the tool comparisons never do: where is your business actually standing today?
This post is the block-by-block record of the session — what each of the seven teaching blocks covers, what happens in the room, and the complete set of 12 participant prompt cards, ready to copy.
Best Answer
WSS-04 is a free 90-minute session built around five stated learning outcomes, seven teaching blocks and one instrument: the Kitchen Check, five questions scored 0 to 4 on Access, Volume, Checking, Initiation and Memory. Your level is your lowest score, and your lowest score is your next project. The session walks all five adoption levels from The Advisor to The Franchise Owner, names the bottleneck and the guardrail at each one, and sends every participant away with 12 prompt cards, 26 copyable prompts and one 30-day move.
The question that decides whether money spent on AI turns into anything is almost never which tool you bought.
“We pay for the business plan and ten of my staff use it every day. So why does nothing more come out the other end?”
That is the question owners actually ask, and the answer is usually arithmetic rather than technology. One step in the workflow never scaled, so nothing downstream of it could. Ten people chatting is ten people at Level 1, not one business at Level 2. WSS-04 exists to find that step in four minutes, name it out loud, and send you home with one thing to fix.
Who This Session Is For
SME owners, directors and department managers — non-technical. The stated assumption about prior knowledge is short: Has used ChatGPT / Claude / Gemini / Copilot in a chat box. Nothing more.
More practically — if two or more of these are true, this is the ninety minutes:
- You pay for a business AI plan and you could not say, out loud, what changed because of it.
- Your staff use AI every day and your throughput looks the same as it did last year.
- One person in your company is racing ahead with AI and nobody else has moved.
- Your prompts have got longer and longer, and the answers stopped getting better.
- You have automated one thing already and you are not sure what the next thing should be.
- You want to know what to fix next, not which tool to buy.
This is not a tools comparison, and it is not a prompting course.
No product is recommended, no plan is compared and no pricing is discussed. The Kitchen Check contains no question about your subscription, deliberately — level is behaviour, not billing. Prompting appears in one block of the seven, as the engine for climbing, not as the subject.
Before You Join
There is nothing to install and nothing to buy. There are two things to bring.
Bring these two things:
- One AI request from your week that disappointed you. You rebuild it live in Block 4. Without one of your own, the exercise becomes a demonstration you watch instead of a rewrite you do.
- Something to write on. You score yourself on five questions during the session and you need those five numbers for everything that follows.
Before you paste anything confidential, here or in the session: follow your own company AI policy and Malaysia’s PDPA obligations. Start on non-sensitive data — price lists and product specifications before customer records, draft emails before signed contracts. Every sample block in the card set on this page uses fictional data.
The 90 Minutes at a Glance
Seven teaching blocks, in the order a decision actually gets made: unsettle the assumption, install the map, score yourself against it, find out what is holding you, get the engine, see it in a real business, then commit to one thing.
Provoke
Chatbot or agent
Chatbot or agent
Ladder
Five kitchens
Five kitchens
Score
The Kitchen Check
The Kitchen Check
Diagnose
Bottleneck, guardrail
Bottleneck, guardrail
Climb
Context beats prompts
Context beats prompts
Commit
One 30-day move
One 30-day move
| Block | What it covers | Minutes | What you leave the block holding |
|---|---|---|---|
| 0 | Open & the provocation — “Are you still in the kitchen, covered in flour?” Chatbot vs agent. | 5 | One distinction you cannot un-hear, and your own last AI task written down. |
| 1 | Act 1 — The Pizza Shop. Walk Levels 0→4 with the analogy, then the SME translation table. | 13 | The five levels in one table, and a rough sense of which kitchen you are standing in. |
| 2 | The Kitchen Check. Live five-question self-assessment. Scoring rule revealed after they score. | 10 | Five honest scores, one lowest letter, and your level. |
| 3 | Act 2 — Why you’re stuck. Bottleneck and guardrail at each level. The four misreads. Cherny’s test. | 10 | The named bottleneck for your row, and the guardrail that removes it. |
| 4 | Act 3 — The climb. Prompt vs context · GCSCO · Delete-Before-You-Add · ARMS · progressive disclosure. Two timed exercises. | 21 | One weak request of your own, rebuilt — and a technique for cutting the next one down. |
| 5 | Act 4 — One level up. U1, U2 and U5 shown at your level vs next level, then the Morning Brief live demo. | 10 | Three jobs you already do, shown one rung higher, and a routine you watched run. |
| 6 | Commit & close. Lowest letter → 30-day move → My AI Context template → three takeaways. | 6 | One sentence, typed: “In 30 days, I will ______.” |
Counted from the run sheet: the seven teaching blocks total 75 minutes and three Q&A windows take 15, which is the full 90. Inside the content blocks there are eight interaction points — six polls, four chat moments and one thumbs-up round — roughly one every eight minutes across the first hour. The session design targets a 70/20 content-to-Q&A split; the deck as built runs 75/15, and there is a documented cut that restores the target.
The Five Learning Outcomes
These five are reproduced as they were signed off, and every one of them has a named piece of evidence attached to it — which is the part most free webinars leave vague. Signed off
| # | Outcome | Block | Evidenced by |
|---|---|---|---|
| LO1 | Distinguish an AI chatbot from an AI agent, and explain in their own business terms why writing longer prompts stops producing gains. | 0, 1 | Opening poll and chat responses |
| LO2 | Locate their own organisation on the 5-level adoption ladder by completing the 5-question Kitchen Check, and justify that placement with observed behaviour rather than the software licence they hold. | 2 | Live self-score, lowest letter declared in chat |
| LO3 | Identify the single bottleneck holding them at their current level, and name the specific guardrail that removes it. | 3, 5 | “My lowest score is my next project” statement |
| LO4 | Apply the Goal · Context · Source · Constraints · Output frame to rewrite one weak prompt from their own week, and apply Delete-Before-You-Add to a bloated one. | 4 | Two timed rewrite exercises, two or three read aloud |
| LO5 | Draft a one-page My AI Context file and commit to one 30-day next-level move. | 6 | Commitment typed in chat, plus the take-home template |
Two of the five carry the session. LO2 and LO3 are the reason it exists — see where you are, then decide what to improve. LO1 is the on-ramp and LO4 and LO5 are the pay-off.
The Five Levels, in One Table
The framework comes from Boris Cherny, the creator of Claude Code, published in July 2026 and written for software teams. The level numbers and the underlying definitions are unchanged; the vocabulary is translated for a business that does not employ engineers. Note where it starts. Level 1 is not the floor.
| Lvl | Name | You are the | Who does the work | Who checks it |
|---|---|---|---|---|
| 0 | The Advisor | Cook | You | You |
| 1 | The Helper | Supervisor | AI — one task | You, every line |
| 2 | The Crew | Boss | AI — several at once | You, the finished output |
| 3 | The Operations Manager | Director | AI — some of it starts itself | The system checks; you review exceptions |
| 4 | The Franchise Owner | Owner | AI — mostly self-started | Different rules for different work |
Read the last two columns down the page. That is the entire framework — everything else is decoration.
The bottleneck at each level, and the guardrail that unblocks it
Cherny’s central finding is that tokens, seats and a bigger plan do not move you up a level. Removing one bottleneck and building one guardrail does. Signed off
| Lvl | What is really blocking you | The guardrail that unblocks it |
|---|---|---|
| 0 | Approval, fear, and cost-per-token thinking. AI can’t reach your files. | A one-page “what AI may and may not touch” policy + one approved tool + start on non-sensitive data |
| 1 | Your attention. You believe you must read everything, so work stays synchronous — you watch instead of starting the next thing. | A definition of done the AI can check itself against: a template, a checklist, a worked sample |
| 2 | Reviewing the output. Six streams of work instead of one. | Automated self-checks before you see it; same quality bar for AI work as for staff work; pre-approved safe actions |
| 3 | Trust, and decision throughput. The tree of work is too deep to babysit. Scaling volume before the loop earns trust is the trap. | Encoded standards (Skills/SOPs), memory, scoped routines on well-defined streams, cost monitoring |
| 4 | Finding the next thing worth automating, and cost at scale. | Per-workflow cost caps, exception-based monitoring, and separate lanes: automated work vs human-gated work |
“Prompting gets you through Level 1. Context gets you to Level 3. Only guardrails get you to Level 4.”
One honest caveat carried from the session design: the agent counts attached to these levels in the original are enterprise software-engineering figures. They describe the shape of the work, not a target. An SME at genuine Level 3 might run six routines, not a hundred agents.
Block 0 · 5 minutes
Open and the Provocation
The one idea: a chatbot answers your question; an agent does the work — and most SMEs are already paying for an agent and using it as a chatbot.
What this block covers Signed off
| Purpose | Disturb the comfortable assumption |
| Outcome served | LO1 |
| Interaction points |
|
What actually happens in the room
- Housekeeping takes thirty seconds. The trainer states the track record once — ERP and automation for Malaysian SMEs since 2008 — and never returns to it. Credibility repeated becomes selling.
- The poll runs before any teaching. Four options, forty-five seconds, and the result is deliberately not discussed. It is saved and replayed in Block 6 so the room can watch its own answer move.
- One distinction is set down, and everything else stands on it. A chatbot answers your question and you still do the work. An agent uses your tools and your files and completes the task.
- The uncomfortable line lands early, not at the end. Most people in the room are already paying for an agent and using it as a chatbot.
- Everyone types their last AI task into the chat. One line, however small. The trainer copies two or three into a notepad and quotes them back later, unnamed, when the diagnosis needs a real example instead of a hypothetical one.
- Nothing is corrected yet. No task is called good or bad. Those lines are evidence for Block 3, not material for Block 0.
“Typing into a box and getting text back is not the revolution. It is the doorway.”
Block 1 · 13 minutes
The Pizza Shop: Five Kitchens
The one idea: same shop, same owner, same demand — the only thing that changes across the five levels is who does the work and who checks it.
What this block covers Signed off
| Purpose | Install the ladder |
| Outcome served | LO1 and LO2 |
| Interaction points | None — this is the one stretch of continuous delivery in the session, and it is kept to thirteen minutes for that reason. |
What actually happens in the room
- Five kitchens, roughly ninety seconds each. One pizza shop, one viewpoint, and only the staffing changes — plus where you are standing in each picture.
- Level 0 is named out loud and early. A Michelin-star chef locked in a room, sliding advice through a mail slot, while you still knead every dough. Being told you are at Level 0 twenty minutes later must not feel like failure, so the floor is set now.
- The analogy is dropped on purpose. It is a delivery vehicle and it is left behind before the scoring starts. A room that enjoys the pictures and learns nothing transferable is the failure mode this block guards against.
- One slide gets five seconds of silence. The ladder table is the screenshot slide, and the instruction is explicit: if you capture one thing today, capture this.
- Expectations are set before anyone is measured. Most businesses in the room are at Level 0 or Level 1, and that is normal rather than slow.
- The target is named, and it is not Level 4. Nobody skips a rung. The goal is plus one.
“An SME at genuine Level 3 might run six routines, not a hundred agents. It is the shape of the work that matters, not the headcount.”
Block 2 · 10 minutes
The Kitchen Check
The one idea: your level is your lowest score, not your average — and your lowest score is your next project.
What this block covers Signed off
| Purpose | The self-location moment |
| Outcome served | LO2 |
| Interaction points |
|
The instrument itself Signed off
Five questions, each scored 0 to 4. It takes four minutes and needs no technical vocabulary. Answer for what actually happens in your business this month — not what you plan, and not what you have paid for.
| The question | How to score it, 0 to 4 |
|---|---|
| A — ACCESS Can AI reach your actual business files and systems? |
|
| V — VOLUME How many AI tasks run for you at the same time? |
|
| C — CHECKING Who checks the work, and how? |
|
| I — INITIATION Who decides that a piece of work should start? |
|
| M — MEMORY & SOP How often do you re-explain your business to the AI? |
|
“Your level is your lowest score, not your average. And your lowest score is your next project.”
The bonus reading is the spread. Take your highest letter minus your lowest. A gap of two or more usually means one person in the company is racing ahead while the business stands still — a risk, not an achievement, because when that person leaves the capability leaves with them.
What actually happens in the room
- The score is pre-framed before a single question is asked. Three sentences: most of you will score lower than you expect and that is the instrument working; answer for this month; nobody sees your numbers except you. A room that feels judged stops typing, and a webinar with a silent chat is a video.
- The five questions run one at a time as live polls, about a minute each, with each result shown before moving on. The room seeing its own shape is half the teaching.
- The scoring rule is withheld until the scores already exist. Only then: do not add them up, do not average them, find your smallest number.
- Then three full seconds of silence. People need the pause to recount their own numbers, and filling it is the most common way this session gets diluted.
- The rule is proved with the room’s own poll results, not argued. Volume 2 and Checking 1 are usually both on screen. If AI drafts eight quotations this afternoon and you read every line of all eight, how many leave the building today? The same number as if it had drafted one.
- Everyone posts one letter — their lowest. The trainer tallies roughly, names the two dominant letters and reads the room back to itself.
- The block closes on a reframe, not a verdict. The rule sounds harsh and is the kindest thing in the framework: you have a one-part problem, and it just named itself.
Block 3 · 10 minutes
Why You Are Stuck
The one idea: a guardrail is not a restriction — it is what lets you let go safely.
What this block covers Signed off
| Purpose | Turn score into diagnosis |
| Outcome served | LO3 |
| Interaction point | Thumbs up or down — four rapid scenarios against one question: “Would a staff member have done this?” |
Four things people say that are not true Signed off
| What people say | What is actually true |
|---|---|
| “We pay for the Pro or Enterprise plan, so we are Level 3.” | Level is behaviour, not billing. One supervised session is Level 1 no matter what you pay. |
| “Ten of my staff use ChatGPT daily, so we are Level 2.” | Ten people chatting is ten Level 1s, not one Level 2. Level 2 needs self-checking work, not more people. |
| “We switched on automation, so we are Level 4.” | Automation without verification and cost control is Level 2 with a bigger bill. |
| “Our AI champion is ten times more productive, so we are advanced.” | That is the problem, not the win. One hero without guardrails widens the gap inside your own company. |
The one question that decides what to automate
“Would one of my staff have done this task, this way?” Yes → automate it. No → keep a human gate on it.
Four scenarios are put to the room, thumbs up to automate and thumbs down to gate. Ninety seconds, total.
| Scenario | The answer, and why |
|---|---|
| Chasing an invoice that is seven days overdue | Up. Your clerk would send exactly that email. |
| Deciding whether to give a customer a RM2,000 refund | Down. That is a judgement about a customer relationship and money. |
| Turning this month’s sales export into a variance summary | Up — with one condition. It must show you which figures it could not tie. |
| Replying to a complaint that mentions a lawyer | Down, hard. And notice: that is a rule you write once, not a technology decision. |
What actually happens in the room
- The word “guardrail” is corrected before it is used. It is not a restriction. Cars go fast because they have brakes; a business with no way to check AI’s work cannot afford to let AI do very much of it.
- Twenty seconds of silence on the bottleneck table. Find your row. It is the only row that matters to you today.
- Level 1 gets a full minute of its own, because statistically most of the room lives there. The bottleneck is your attention, not AI’s ability — and watching is not working.
- The fix is a document, not a decision. “Trust it more” is not on offer; trust is an outcome. What is on offer is a definition of done the AI can check itself against: a template, a checklist, or one approved worked example.
- One of the tasks typed in Block 0 is read back, unnamed. For that task, what would “done and correct” look like written down? Three bullets is a first guardrail.
- The four misreads are delivered fast, almost as a list. The rhythm is the point — each one is a sentence people have said to the trainer, answered in a sentence.
- The block ends on a decision rule, not a warning. From Level 3 onward, one question decides everything, and it needs no technical knowledge to apply.
Block 4 · 21 minutes
The Climb: Context Beats Prompts
The one idea: every bottleneck on the ladder is either a context problem or a checking problem — and neither one is fixed by a longer prompt.
What this block covers Signed off
| Purpose | Give them the engine |
| Outcome served | LO4 |
| Interaction points |
|
Five things to settle before any important request Signed off
| Ask yourself | Worked example | |
|---|---|---|
| Goal | What do I want produced? | “A one-page summary for management.” |
| Context | What must AI understand? | “The audience is the board. They care about cash, risk and next steps.” |
| Source | What should it work from? | “Use the attached Q2 sales report.” |
| Constraints | What really matters, and what must not happen? | “Do not invent numbers. Flag anything missing.” |
| Output | What should it look like? | “Headings, short paragraphs, one action table.” |
This is not a longer prompt. It is a better-organised one, and it is usually shorter than what you write today. The most-skipped of the five is Source: give AI the document instead of describing the document, because the real file contains the details you would have forgotten to mention.
Delete before you add
When the answer is poor, the instinct is to add a rule. The signature technique of the session is to do the opposite first, in five passes:
- Delete instructions that repeat each other.
“Be professional” written four times is one instruction and three wasted lines. - Delete instructions that contradict each other.
“Be detailed” and “keep it short” cannot both win. Pick one. - Delete descriptions of things AI can already see
in the file you attached. - Delete rules for things AI can reasonably decide by itself.
Paragraph counts. Sentence lengths. Which greeting to use. - Keep only the constraints that genuinely matter
— the ones where being wrong costs you money, a customer, or a legal problem. Then try again.
Level 2 to Level 3 is not a prompting move Signed off
It is a workspace move. Build ARMS — and notice the right-hand column, because ARMS is not a new topic. It is the direct fix for the letters you scored yourself on twenty minutes earlier.
| What it means for an SME | Which letter it fixes | ||
|---|---|---|---|
| A | Applications | The tools and files AI is actually connected to | A |
| R | Routines | Work that runs on a schedule or a trigger, without you | I |
| M | Memory | Standing context — your My AI Context file, past reports, Projects | M |
| S | Skills | Your SOPs written down once, reused every time | M and C |
What actually happens in the room
- The bridge is made by pointing backwards. Look at every row of the bottleneck table: each one is a context problem or a checking problem. Nothing else is on that list.
- Prompt and context are pulled apart. A prompt is what I want now; context is what AI needs to understand the situation. You do not hand a new manager fifty rules — you hand them the file and tell them what matters.
- Four contradictory instructions are read off one line to show why longer prompts often get worse: part of the effort now goes on deciding which rule you meant.
- Four silent minutes. Participants rebuild their own disappointing request using the five. The trainer stays quiet for the first two minutes — no narrating, no encouraging — then offers a floor: Goal and Source alone is already most of the value.
- Two rewrites are read aloud. One strong, named for what makes it strong. One that described its source instead of attaching it — and that person is never named.
- A 140-word bloated prompt gets cut in the chat, and the real question lands underneath it: which of these rules would you actually give a new staff member?
- ARMS closes the block, mapped straight back to the letters, along with progressive disclosure: point AI at the one or two documents the job needs, not the whole handbook.
“Most bad AI results are a context problem wearing a prompt costume.”
Block 5 · 10 minutes
One Level Up, in a Real SME
The one idea: same job, one rung higher — and the owner’s job changes from improving the output to improving the rules.
What this block covers Signed off
| Purpose | Make the next rung concrete |
| Outcome served | LO3 and LO5 |
| Set piece | A four-minute live demonstration of a scheduled routine and the brief it produced that morning, with a recorded fallback ready. |
The same job, walked up all five rungs
Six jobs are named; three are walked. This is the first of them — the task the trainer describes as the most common “I do this fifteen times a week” job in a Malaysian SME. Signed off
| Lvl | What the quotation desk looks like |
|---|---|
| 0 | Owner asks a chatbot “how should I word a quotation?”, then types it into Excel himself. |
| 1 | Uploads the RFQ email and the price list. AI drafts the quote. Owner reads every line and retypes it into the company template. |
| 2 | AI works inside the quotation folder: reads the RFQ, pulls the price list, drafts into the company template, checks itself against a margin rule and a no-missing-field checklist. Eight quotes ready; owner reviews finished PDFs, not workings. |
| 3 | A routine watches the sales inbox. New RFQ → quote drafted, margin-checked, filed, and a Telegram message goes to the sales manager: ready to send or needs your decision. The owner now improves the pricing rules, not the quotes. |
| 4 | Quotes below a value threshold and above a margin threshold go out on their own. Anything outside the bounds is routed to a human. The owner sets the bounds and reads the exception report. |
What actually happens in the room
- Six jobs are named, three are walked. All six live in the card set below; the Quotation Desk, the Collections Clerk and the Morning Brief get the live treatment.
- The Quotation Desk lands on one line from its card, not on the automation: never send the quotation to the customer — only the sales manager sends. A Level 3 routine is defined as much by what it must not do as by what it does.
- The Collections Clerk is used to show the checking column moving. At Level 1 you check every letter. By Level 3 you check the rules, not the letters.
- Then the demonstration, in a fixed order. The output first, with its timestamp pointed at. The routine second — a paragraph of plain English on a timer. The guardrail third.
- The guardrail line gets its own moment: “If there is nothing worth reporting, say nothing significant — do not pad it.” Without that line a daily routine manufactures importance every morning and stops being read inside two weeks. Most Level 3 routines fail there, not on the technology.
- A recorded capture sits open in a second tab. If the live version fails, it switches in one sentence and keeps talking. No troubleshooting on air.
- The block closes by naming three letters, not a prompt. It started itself — that is I. It used real sources — that is A. It checked itself against a rule — that is C. At Level 3, prompting is the least important thing you do.
Block 6 · 6 minutes
Commit and Close
The one idea: one move, not five — your lowest letter already chose it for you.
What this block covers Signed off
| Purpose | Convert insight to action |
| Outcome served | LO5 |
| Interaction point | Chat — “Type your one 30-day move.” The chat is screenshotted, and it becomes the follow-up list. |
Your lowest letter, and the one move it chooses
| Lowest letter | Your 30-day move | Which card to use |
|---|---|---|
| A — Access | Connect AI to one real folder or inbox this week. Non-sensitive data first. Write the one-page “what AI may and may not touch” policy before you do. | Any card, Level 2 prompt |
| V — Volume | Run two tasks at once instead of one — and stop watching the first one while the second runs. | U1 and U3, Level 2 prompts |
| C — Checking | Write one checklist that defines “done and correct” for your most repeated task. Then add a check report line to your prompt. | U1 or U6, Level 2 prompt |
| I — Initiation | Take the task you do every single week and put it on a schedule. | U2 or U5, Level 3 setup |
| M — Memory | Write your one-page My AI Context file today, then turn one tribal process into a written procedure. | Card M, then U7 |
What actually happens in the room
- The opening poll comes back on screen, next to the room’s lowest letters. Most people move down a level after the Kitchen Check, and that is the instrument doing its job rather than anyone going backwards.
- One move each, chosen by the lowest letter, read straight off a five-row table. Nobody chooses from a menu of five improvements.
- The commitment is typed as a single sentence — “In 30 days, I will ______” — and the chat is captured. Three are read aloud.
- The reason for only one move is given as arithmetic, not encouragement. Four of your five letters are already ahead of the lowest one; working on them builds capability you cannot yet use, which you pay for and do not collect.
- The My AI Context file gets exactly one rule: keep it to one page. A thirty-page context file is a Level 1 habit wearing a Level 3 costume — the same over-instruction problem, stored somewhere else.
- Three takeaways close it, and one question to keep asking from Level 3 onward.
“An honest Level 1 will overtake an optimistic Level 3 every single time — because the honest one knows what to fix.”
What Every Participant Carries Out
Six things, and five of them are produced by the participant during the ninety minutes rather than handed over at the end.
1. Five scores and one letter
Your Kitchen Check numbers for Access, Volume, Checking, Initiation and Memory — and the lowest one, which is your level.
Block 2 · produced live
2. A named bottleneck
The single thing holding you at your current level, taken off the bottleneck table, plus the specific guardrail that removes it.
Block 3 · produced live
3. One request, rebuilt
A disappointing request from your own week, restructured with Goal, Context, Source, Constraints and Output — usually shorter than the original.
Block 4 · produced live
4. One 30-day move
A single written sentence — “In 30 days, I will ______” — chosen for you by your lowest letter, not from a menu.
Block 6 · produced live
5. The 12-card prompt pack
The Kitchen Check, two technique cards, seven use-case cards and two take-home cards, with 26 prompts you can copy. All of it is on this page.
Take home
6. The My AI Context template
A one-page file that stops you re-explaining your business every morning. Twenty minutes, once. It is card M below.
Take home
Sent to everyone afterwards, including anyone who registered and could not attend: the deck, the Kitchen Check scorecard as a printable page, and the 30-day planner. The scorecard is built to be re-run every ninety days — one line asks for today’s five numbers, the line under it asks for the same five numbers in ninety days.
The Complete Prompt Card Set
All 12 cards from the participant handout, with 26 copyable prompts between them. One scorecard, two technique cards, seven use-case cards and two take-home cards. Every use-case card shows the same job at each level: Level 0 has no prompt, because at Level 0 you do the work yourself and that is the point, and Level 4 has no prompt either, because at Level 4 you write rules and lanes rather than requests.
Use the level one above where you are. If you scored Level 1, use the Level 2 prompt. Do not jump to Level 3 — a Level 3 routine built on a Level 1 habit just fails faster.
Two things before you paste any of these:
- Replace everything in <angle brackets> and [square brackets] with your own detail. A card used unchanged gives you a generic answer, and you will blame the AI.
- Fictional or sandbox data only. Never point a card at live customer or staff data while you are still testing. The sample blocks below use invented companies, invented people and invented figures. Replace real names, identity-card numbers, bank details and account numbers with placeholders. Follow your own company AI policy and your PDPA obligations.
Where the work is technical, a professional still signs it.
Several of these cards touch collections, month-end reporting and customer complaints that may become legal matters. AI output is a first draft. Anything that will be filed, published or relied on — a tax treatment, a provision, a judgement call, a response to a legal threat — is reviewed by a qualified professional before it leaves your business. Cards U2, U4 and U6 carry that boundary in their own wording.
Click any card to open it. The Copy button on each prompt and the Open-all control come from a small helper script — if your browser or platform strips it, every card still opens and every prompt is still selectable.
Score yourself first 1 card
Find your lowest letter before you touch anything else. Every other card is chosen by that one number.
K The Kitchen Check — score yourself PEN AND PAPER EVERY PARTICIPANT TAKE HOME
When: First, before any other card. Then again every ninety days.
Score yourself 0 to 4 on each of the five letters — the full wording of all five options for each letter is set out in Block 2 above. Answer for what actually happens this month, not what you plan and not what you have paid for.
| A | V | C | I | M | |
|---|---|---|---|---|---|
| Today — date: ________ | |||||
| In 90 days — date: ________ |
What good looks like: Five numbers written down, the lowest one circled, and a date ninety days out already in your calendar to score it again. Your level is your lowest score, not your average — and your lowest score is your next project.
Imbalance check: highest score minus lowest score. A gap of 2 or more usually means one person in your company is racing ahead while the business stands still. That is a risk, not an achievement.
The two techniques 2 cards
One frame for building a request, one discipline for cutting one down. They apply to every card below.
T1 The Five Things — Goal · Context · Source · Constraints · Output ANY AI ASSISTANT EVERY REQUEST THAT MATTERS
When: Before any request that matters. Practised live in Block 4.
It is not a longer prompt — it is a better-organised one, and it is usually shorter than what you write today.
Goal: [what you want produced] Context: [who it is for, what they care about, anything AI cannot work out for itself] Source: [the actual document, file or data to use — and "use only this"] Constraints: [what must not happen · what must not be invented · what you decide, not AI] Output: [format, length, structure]
You do not need all five every time. You need the ones AI cannot reasonably work out for itself. If the goal and the source are obvious, two lines is a complete prompt.
The most-skipped one is Source. Give AI the document instead of describing the document. The real file contains the details you would have forgotten to mention.
T2 Delete Before You Add ANY AI ASSISTANT WHEN THE ANSWER IS POOR
When: The answer came back poor and your instinct is to add another rule.
Run these five passes over your prompt before you add anything.
1
Delete instructions that repeat each other. “Be professional” written four times is one instruction and three wasted lines.
2
Delete instructions that contradict each other. “Be detailed” and “keep it short” cannot both win. Pick one.
3
Delete descriptions of things AI can already see in the file you attached.
4
Delete rules for things AI can reasonably decide by itself. Paragraph counts. Sentence lengths. Which greeting to use.
5
Keep only the constraints that genuinely matter — the ones where being wrong costs you money, a customer, or a legal problem. Then try again.
When it still gets it wrong — the six-way check. Add one thing only.
| What went wrong | What to add — one thing only |
|---|---|
| It misunderstood the task | A clearer Goal |
| It misunderstood the situation | More Context |
| It used wrong information | The real Source |
| It broke something important | One clear Constraint |
| The answer is too complicated | A simpler Output requirement |
| It keeps making the same mistake | Stop adding. Look for the contradiction you have not spotted yet. |
What good looks like: Very often the better prompt is the shorter prompt. Add context because it solves a real problem — not because you can.
The seven use cases 7 cards
Every card shows the same job at each level. Level 0 has no prompt — at Level 0 you do the work yourself, and that is the point. Level 4 has no prompt either — at Level 4 you write rules and lanes, not requests. Use the level one above where you are.
U1 The Quotation Desk ANY AI ASSISTANT TRADING · DISTRIBUTION · MANUFACTURING · SERVICES LIFTS A · C
When: An RFQ arrives by email and someone has to turn it into a priced, checked quotation. Walked live in Block 5.
Level 0 — you read the RFQ, open Excel, and type. No prompt. This is the rung you are leaving.
1
Level 1 — one task, you check every line
Goal: Draft a quotation for the customer request below. Context: We are [company], we sell [products]. This customer is [new / existing since ____]. They care most about [price / lead time / specification]. Source: The RFQ is pasted below. Our price list is attached. Use only these. Constraints: Use only prices from the attached price list. Do not invent stock availability or lead times — if it is not in the source, write "to confirm". Output: Our standard layout — item, quantity, unit price, total, lead time, validity period. [paste the RFQ here]Try it now — ready to run, sample data included. Nothing to attach, nothing to fill in. Copy this whole block and paste it straight into your AI.
Goal: Draft a quotation for the customer request below. Context: We are Metalworks Supply Sdn Bhd. We sell stainless and mild steel fabricated furniture for warehouses, kitchens and clinics. This customer is new. They care most about price and lead time. Source: The RFQ and our price list are both pasted below. Use only these. Constraints: Use only prices from the price list. Do not invent stock availability or lead times - if it is not in the source, write "to confirm". Output: Our standard layout - item, quantity, unit price, total, lead time, validity period. --- RFQ, received by email --- From: Faridah Osman <purchasing@sinarlogistics.com.my> Subject: RFQ - stainless worktables and shelving, Shah Alam warehouse Hi, We are fitting out a new packing area at our Shah Alam warehouse and need a quotation for the following: 1. Stainless steel worktable, 1800 x 750mm, with lower shelf - 6 units 2. Stainless steel worktable, 1200 x 600mm - 4 units 3. Wall-mounted shelf, stainless, 1500mm - 8 units 4. Mobile trolley, 3 tier, 900 x 600mm - 2 units Please include delivery to Shah Alam, and installation if available. We want to confirm by end of this month, so please advise your lead time. Also, do you have a mild steel powder-coated option? Budget is tight this quarter. Thanks, Faridah Osman Purchasing Executive, Sinar Logistics Sdn Bhd --- OUR PRICE LIST, effective 1 Jan 2026, RM, excluding SST --- SS-WT18 Worktable, stainless, 1800x750mm, lower shelf 1,280 2 weeks SS-WT12 Worktable, stainless, 1200x600mm 860 2 weeks SS-SH15 Wall shelf, stainless, 1500mm 340 in stock SS-SH12 Wall shelf, stainless, 1200mm 295 in stock MS-WT18 Worktable, mild steel powder-coated, 1800x750mm 790 4 weeks DEL-KV Delivery within Klang Valley 250 -
What good looks like: The mobile trolley and installation are not on the price list, and neither is a 1200mm mild steel table. A correct answer writes “to confirm” for all three instead of guessing. If your AI invents a price for the trolley, you have just watched a constraint fail.
2
Level 2 — real sources, and it checks itself before you see it
Goal: Produce a ready-to-send quotation from the RFQ in this folder. Context: The customer is [name]. Their history is in [file]. Standard payment terms are [terms]. Source: Use the RFQ, the current price list and the quotation template — all in this folder. Do not use anything else. Constraints: - Every line item must trace to a row in the price list. - Gross margin must be at least [__]%. If any line falls below, stop and tell me instead of quoting. - Never state a lead time that is not in the stock file. Output: The completed quotation in our template, plus a three-line check report: (1) all fields filled, yes or no (2) the lowest margin on this quote (3) anything you had to mark "to confirm".
The three-line check report is the whole Level 2 move. You stop reading the working and start reading the check. That is what lets you review eight quotes in the time you used to review one.
3
Level 3 — write it once, it runs without you
Run this whenever a new email arrives in [sales inbox] whose subject or body asks for a price, a quotation or an RFQ. 1. Read the request and identify the customer from our customer list. 2. Draft the quotation using the price list and the quotation template. 3. Check it against the quotation rules file. 4. Save the draft to [folder] as [customer]-[date]. 5. Send a Telegram message to the sales manager with: customer name, total value, the lowest margin on the quote, and one word — READY if every check passed, or DECISION if a check failed, naming which one. Never send the quotation to the customer. Only the sales manager sends.
Read that last line twice. A Level 3 routine is defined as much by what it must not do as by what it does.
4
Level 4 — rules and lanes, not a prompt
- Automatic lane: existing customer · total under RM[____] · every margin above [__]% · standard items only · standard payment terms. Goes out without a human.
- Human lane: new customer · above the value threshold · any non-standard item · any margin below the floor · any request for extended terms.
- Cost cap: limit the routine to [__] runs a day and alert if exceeded.
- Exception report: a daily list of what went out on its own, and what was routed to a person.
Before you use this card: fill in your margin floor, your value threshold, your folder names and who the approver is. Those four numbers are the entire difference between a useful routine and a dangerous one.
U2 The Collections Clerk ANY AI ASSISTANT EVERY SME WITH INVOICES LIFTS C · I
When: Overdue invoices need chasing, weekly, in a tone that gets paid without losing the customer. Walked live in Block 5.
1
Level 1
Goal: Draft payment reminder emails for the overdue accounts below. Context: We are [company]. These are business customers we want to keep. Tone: firm and warm. Never threatening. Source: The aged debtors list is pasted below. Use only the figures in it. Constraints: Never state a figure that is not in the list. Do not mention legal action. Do not offer a discount or a payment plan — that is my decision, not yours. Output: One short email per customer with a subject line, grouped into 1-30 days, 31-60 days and over 60 days, with a slightly firmer tone at each stage. [paste the aged debtors list here]Try it now — ready to run, sample data included. Nothing to attach, nothing to fill in.
Goal: Draft payment reminder emails for the overdue accounts below. Context: We are Metalworks Supply Sdn Bhd. These are business customers we want to keep. Tone: firm and warm. Never threatening. Source: The aged debtors list is pasted below. Use only the figures in it. Constraints: Never state a figure that is not in the list. Do not mention legal action. Do not offer a discount or a payment plan - that is my decision, not yours. Anyone flagged "dispute" gets no reminder - list them separately for me instead. Output: One short email per customer with a subject line, grouped into 1-30 days, 31-60 days and over 60 days, with a slightly firmer tone at each stage. --- AGED DEBTORS as at 30 September 2026, RM --- Customer Invoice Amount Days over Notes Teguh Manufacturing Sdn Bhd INV-2841 12,400 8 - Kedai Runcit Aman INV-2799 1,860 22 - Prisma Interiors Sdn Bhd INV-2755 8,300 41 Dispute - claims short delivery Wawasan Trading INV-2702 24,750 63 - Lim & Sons Hardware INV-2688 3,120 71 Promised payment 15 Oct Bayu Cafe Group INV-2610 19,900 96 -
What good looks like: Prisma Interiors is flagged dispute — a correct answer sends them nothing and lists them separately. Watch also whether it invents a payment plan for Bayu Cafe at 96 days. It was told not to.
2
Level 2
Goal: Prepare this week's collection pack from the aged debtors export in this folder. Context: Customer relationship notes are in [file]. Our payment terms are [terms]. Source: The aged debtors export and the customer notes in this folder. Nothing else. Constraints: - Every amount and invoice number must come from the export. If a figure is missing, write "please confirm" rather than guessing. - Never mention legal action. Never offer a discount or a payment plan. - Anyone flagged "dispute" in the notes gets no reminder — list them separately. Output: (1) A folder of draft emails, one per customer, grouped by ageing bucket. (2) A one-page summary: total overdue, the five largest, who is in dispute, and any figure you could not verify against the export.
3
Level 3 setup
Every Monday at 8am: 1. Pull the aged debtors report from [Odoo / accounting system]. 2. Accounts 1-30 days overdue with no dispute flag: send the gentle first reminder from the template and log it in the customer record. 3. Accounts 31-60 days: draft the reminder but do not send. Put it in the account manager's folder with a suggested call script. 4. Accounts over 60 days, in dispute, or over RM[____]: draft nothing. Add to the escalation list. 5. Send me one message: how many reminders went out, total overdue, and the escalation list. Never send to an account flagged "dispute". Never send the same customer two reminders in the same week.
4
Level 4 — rules and lanes
- A full reminder ladder runs per customer risk tier, set once.
- Only disputes and accounts over 90 days reach a person.
- Tone audit monthly: a sample of sent reminders reviewed against the tone rules.
- Hard stop: no automated message ever goes to a customer with an open dispute or an open credit note.
Before you use this card: decide your escalation ringgit threshold and who owns the 31-60 day bucket. If nobody owns it today, the routine will simply automate an unowned queue.
U3 The Content Kitchen ANY AI ASSISTANT RETAIL · F&B · SERVICES LIFTS M · C
When: One product or promotion has to become four different pieces of content, on brand, without inventing claims.
1
Level 1
Goal: Write a Facebook post, a LinkedIn post and a WhatsApp broadcast for the product below. Context: We sell to [audience] in [market]. Our tone is [tone]. The key message is [message]. Source: The product details are pasted below. Use only what is there. Constraints: Do not claim anything about performance, certification or price that is not in the source. Keep the WhatsApp version under 80 words. Output: Three separate posts, clearly labelled, each ready to paste. [paste the product details here]Try it now — ready to run, sample data included. Nothing to attach, nothing to fill in.
Goal: Write a Facebook post, a LinkedIn post and a WhatsApp broadcast for the product below. Context: We sell to homeowners and small offices in Klang Valley. Our tone is friendly, practical, no hype. The key message is that damp season is coming. Source: The product details are pasted below. Use only what is there. Constraints: Do not claim anything about performance, health benefits, certification or price that is not in the source. Keep the WhatsApp version under 80 words. Output: Three separate posts, clearly labelled, each ready to paste. --- PRODUCT SPEC SHEET --- Product: TeraDry 20L Dehumidifier, model TD-20 Capacity: 20 litres per day at 30 degrees C, 80% relative humidity Tank: 4.5 litres, auto shut-off when full Coverage: up to 60 square metres Power: 320W, 240V Noise level: 42 dB Filter: washable pre-filter Warranty: 2 years parts, 1 year compressor Colour: white Weight: 12.5 kg Retail price: RM 899 SIRIM certified
What good looks like: The spec sheet says nothing about energy saving, killing mould, or being the best in Malaysia. If any of those appear in the posts, the AI invented a claim — and in a real campaign that is your liability, not its.
2
Level 2
Goal: Produce the full launch pack for the product in this folder. Context: Our brand guide is in this folder — our voice, the words we use, the words we never use, and how we talk about price. Source: The product spec sheet, the brand guide, and two past posts that worked well. All in this folder. Constraints: - Follow the brand guide, not the past posts. The past posts show tone; they are not templates to copy. - No claim may appear that is not in the spec sheet. - The Malay and English versions must carry the same claims. Do not add a claim in one language only. Output: Facebook post, LinkedIn post, WhatsApp broadcast, and a one-page product sheet. Then a short check list: every claim you used, and which line of the spec sheet it came from.
“Which line did that claim come from?” is a marketing self-check and a compliance record at the same time. It is the cheapest protection an SME can build.
3
Level 3 setup
Whenever a new product folder appears in [folder]: 1. Read the spec sheet and the brand guide. 2. Produce the launch pack: Facebook, LinkedIn, WhatsApp, product sheet. 3. Check every claim against the spec sheet and list the source of each one. 4. Place the pack in the approval queue and notify [person] on Telegram with the product name and any claim you could not source. Do not publish anything. Approval is always human.
4
Level 4 — rules and lanes
- Performance data feeds back; the system proposes the next variation within brand and budget limits.
- Claims stay human-gated permanently. Tone and format can be automated; a factual claim about your product cannot.
- Budget cap per campaign, set once.
U4 The Complaint Counter ANY AI ASSISTANT F&B · RETAIL · SERVICES · LOGISTICS LIFTS C · A
When: Messages arrive all day and someone has to work out what each one needs, reply to the easy ones, and get the serious ones to a human fast.
1
Level 1
Goal: Draft a reply to the customer message below. Context: We are [company]. This customer [is new / has ordered __ times]. Our position on [refunds / replacements] is [policy]. Source: The message is pasted below and the order details are attached. Constraints: Do not promise a refund, a replacement, a discount or a delivery date. If the message mentions injury, safety, a lawyer or the authorities, stop and tell me instead of drafting a reply. Output: A short reply in [language], plus one line telling me what the customer actually wants. [paste the message here]
Look at that constraint again. “Stop and tell me instead of drafting” is a guardrail written as a sentence. You do not need software to build your first guardrail — you need one clear line.
Try it now — ready to run, sample data included. Nothing to attach, nothing to fill in.
Goal: Draft a reply to each customer message below, and tell me what each customer actually wants. Context: We are Metalworks Supply Sdn Bhd. Our position on refunds is that a manager approves every one, case by case. Source: The two messages are pasted below. Constraints: Do not promise a refund, a replacement, a discount or a delivery date. If a message mentions injury, safety, a lawyer or the authorities, stop and tell me instead of drafting a reply. Output: For each message - what the customer wants in one line, then either a short reply, or a clear statement that you are not drafting one and why. --- MESSAGE 1, WhatsApp --- Hi, I ordered 2 sets last Tuesday, order SG-4471, and only 1 arrived today. The box was also wet on one corner. I already called your office twice and nobody called me back. I need the second set before Friday because I have an event. If not I want a full refund. Very disappointed with your service. --- MESSAGE 2, WhatsApp --- Your product cut my staff's hand when she was unpacking it this morning. There was a sharp metal edge on the frame. We had to send her to the clinic for stitches. I am consulting my lawyer about this. I want compensation.
What good looks like: Message 1 should get a reply with no refund promised and no date given. Message 2 should get no draft at all — it mentions injury and a lawyer, so the correct answer is to stop and escalate. One prompt, two opposite behaviours, from a single constraint line.
2
Level 2
Goal: Triage today's customer messages in this folder and prepare the replies. Context: Our service policy and reply templates are in this folder. Order history is in [file]. Source: The message export, the policy and the order history. Nothing else. Constraints: - Classify each message: routine question / delivery issue / product fault / refund request / escalate. - Draft replies only for routine questions and delivery issues. - Escalate immediately with no draft: anything mentioning injury, safety, legal action or the authorities, and any customer who has complained more than twice this month. - Never promise money, a replacement or a date. Output: A table — customer, category, what they want, draft ready yes or no — with escalations at the top. Plus a folder of the drafts.
3
Level 3 setup
Whenever a message arrives in [WhatsApp Business / Telegram / support inbox]: 1. Classify it using the categories in the service policy. 2. If it matches a question in the FAQ file, reply directly using the FAQ answer and log it. 3. If it does not, open a ticket containing: what the customer wants, their order history, and a suggested resolution. Assign it to [team]. 4. If it mentions injury, safety, legal action, the authorities, or is from a customer who has complained more than twice this month, escalate to [person] on Telegram immediately and reply only with the holding message in the template. Never promise money, a replacement or a delivery date in an automatic reply. Send me a daily summary: how many handled automatically, how many ticketed, how many escalated.
4
Level 4 — rules and lanes
- A resolution pathway per category, each with a spend limit — for example, goodwill vouchers under RM[__] approved automatically, anything above routed to a person.
- Humans see exceptions and trends, not individual routine messages.
- Weekly: the categories that grew get reviewed. A rising category usually means a product or process problem, not a service problem.
Before you use this card: write your escalation word list first — injury, safety, lawyer, saman, KPDN, refund above RM[__]. That list is the guardrail. Everything else is convenience.
U5 The Morning Brief ANY AI ASSISTANT OWNERS AND MDS LIFTS I · A
When: You need to know what happened in your industry and your inbox before your first meeting, without spending an hour on it. Demonstrated live in Block 5.
1
Level 1
Goal: Give me a one-page brief on what happened in [industry] this week. Context: I run [business] in [market]. I care most about [competitor moves / raw material prices / regulation / customer demand]. Source: The five links pasted below. Use only these. Constraints: Say clearly what is fact from the source and what is your interpretation. If something appears in only one source, say so. Output: Five bullets on what happened, three lines on what it means for us, and one line on anything I should act on this week. [paste the links here]
No sample pack for this card. Inventing fake news to practise on would teach you the wrong habit. Use three or four real links you already have open — or skip straight to the Level 2 prompt, which does the searching itself and needs nothing pasted in.
2
Level 2
Goal: Research these four topics and give me one combined brief: [topic 1], [topic 2], [topic 3], [competitor names]. Context: I run [business] in [market]. Decisions I am currently weighing: [decision 1], [decision 2]. Source: Search current sources. Prefer primary sources — company announcements, regulator notices, published figures — over commentary about them. Constraints: - Every claim must carry its source link. - If you cannot find a source for something, leave it out rather than guessing. - Mark clearly which items are new since [date] and which are background. Output: One page. Three sections — What happened · What it means for us · What needs a decision. Then a source list.
3
Level 3 setup — this is the one demonstrated in the session
Every weekday at 5:30am: 1. Search for news from the last 24 hours on: [topic 1], [topic 2], [competitor names]. 2. Scan my inbox for anything from a customer or supplier marked urgent, or containing the words [word list]. 3. Write a brief with three sections: What happened · What it means for us · What needs a decision today. 4. Every claim must carry its source link. If you cannot find a source, leave the item out rather than guessing. 5. Save it to [folder] and send me the link on Telegram by 6:30am. Keep it under one page. If there is nothing worth reporting, say "nothing significant" — do not pad it.
“If there is nothing worth reporting, say nothing significant.” Without that line, a daily routine will manufacture importance every single morning, and you will stop reading it inside two weeks. Most Level 3 routines fail on this, not on the technology.
4
Level 4 — rules and lanes
- The brief triggers work. A competitor price move creates a pricing review task and routes it to the right person.
- Cost cap on research volume per day.
- Monthly review: which items in the brief actually led to a decision? Drop the topics that never do.
U6 The Month-End Pack ANY AI ASSISTANT MANUFACTURING · TRADING · PROFESSIONAL SERVICES LIFTS I · C
When: Every month, the same pack — variance commentary, cash position, and the things the owner should worry about.
1
Level 1
Goal: Write the variance commentary for the month below. Context: We are [business]. Management cares most about [gross margin / overheads / cash]. Source: The trial balance and last month's figures are attached. Use only these. Constraints: Do not invent or estimate any figure. If a number looks wrong, flag it rather than explaining it. Separate what the numbers show from what you think caused it. Output: One page — the five biggest movements, a one-line possible explanation for each clearly marked as an assumption, and a list of anything the accountant needs to confirm.
2
Level 2
Goal: Build the month-end pack from the exports in this folder. Context: Our reporting format is in [file]. Management reviews it on the [__] of each month. Source: The trial balance, the aged debtors and creditors, and the bank summary — all in this folder. Nothing else. Constraints: - Every figure must tie to a source file. List anything you could not tie. - Flag any movement over [__]% or RM[____] against last month. - Do not estimate. "Not available" is an acceptable answer. Output: The pack in our format — variance commentary, cash position, top five risks. Then a reconciliation note: which figures tied, which did not, and where each came from.
3
Level 3 setup
On the [3rd] working day of each month: 1. Pull the exports from [system]. 2. Build the pack in our reporting format. 3. Flag every movement over [__]% or RM[____]. 4. Send the pack to me and to [accountant] with a one-line summary of the largest movement and anything that did not tie. If any required export is missing, do not build a partial pack. Tell me what is missing and stop.
4
Level 4 — rules and lanes
- Continuous close. Anomalies are flagged as they occur, not four weeks later.
- Month-end becomes a review of exceptions that were already handled.
- Anything touching tax treatment, provisions or judgement calls stays human-gated permanently.
Accounting boundary: everything this card produces is a first draft for your accountant or tax agent to review. Nothing here is a filing, and nothing here is tax advice.
U7 The Tribal Knowledge Skill ANY AI ASSISTANT EVERYONE LIFTS M
When: “Only one person here knows how to do this properly.” Turn that into something written down, so AI — and the next new hire — can follow it.
This card works differently from the others. It is a prompt that interviews you.
Goal: Turn the way we do [task] into a written procedure that AI can follow every time. Context: Right now only [person] knows how to do this properly. I am going to describe it, and I will get some of it wrong or leave things out. Source: I will describe the steps. Here are three finished examples of what a good result looks like: [attach three real examples]. Constraints: - Ask me one question at a time. No more than eight questions in total. - Do not write the procedure until you have finished asking. - Where I am vague, ask me for the rule — do not ask me for more examples. Output: A one-page procedure covering: when it applies · the steps in order · the decisions and who makes each one · what "done and correct" looks like · what to do when something is unusual.
“Ask me for the rule, not more examples.” This is the difference between AI copying what you did and AI understanding what you meant. It is also, almost always, the moment an owner discovers their own process has a gap in it.
What good looks like: A one-page procedure you save as a file your AI can read every time. That single file is what moves your M score from 1 to 3 — and M is the letter that unlocks everything above it.
Take home and commit 2 cards
Twenty minutes of writing, and one sentence. These two are the difference between an interesting ninety minutes and a different quarter.
M My AI Context — the one-page file ANY AI ASSISTANT LIFTS M TAKE HOME
When: Today, once. Then you stop re-explaining your business every morning.
Save it as a file your AI can read — or paste it into your tool’s memory, Projects or custom instructions.
MY AI CONTEXT — [your name], [company] Last updated: [date] ABOUT MY WORK [Three lines. What the business does, who we sell to, roughly how big.] MY TYPICAL AUDIENCE [Who I write to. Who actually decides. What they care about.] MY PREFERRED STYLE [How we sound to customers. Formal or plain. Any language mix.] HOW I WANT AI TO HELP [For example: give practical actions before theory. Separate facts from assumptions. Flag risks I have not asked about. Tell me when you are unsure.] IMPORTANT TERMINOLOGY [Words that mean something specific in our company. Product names. What we call our customers. Anything an outsider would get wrong.] IMPORTANT RULES [Only the ones that genuinely matter. Three or four, not thirty.]
What good looks like: One page. Only what AI cannot reasonably work out for itself. You add to it when AI makes the same meaningful mistake twice — not because you thought of something else to say.
One page, and no more. A thirty-page context file is a Level 1 habit wearing a Level 3 costume — the same over-instruction problem, just stored somewhere else.
Keep out of it: customer personal data, identity-card numbers, bank details, staff salaries, and anything your company policy or the PDPA would not allow you to hand to an outside party.
P My 30-Day Next-Level Move PEN AND PAPER EVERY PARTICIPANT TAKE HOME
When: At the end of the session, and again every ninety days. One move, not five — your lowest letter already chose it for you.
The five rows are set out in Block 6 above. Write the commitment down; a move that is not written is a move that does not happen.
My lowest letter today is: ______ My level is: ______ In 30 days, I will _________________________________________________ The one card I will use: ______ I will know it worked when _________________________________________ Date I will re-run the Kitchen Check: ______________
What good looks like: One sentence, one card, one date. Four of your five letters are already ahead of your lowest one — working on them produces capability you cannot yet use. Lift the lowest letter by one, and the four above it start delivering what you already paid for.
What To Do Next
- Score yourself now, before you read anything else again.
Open card K and write down five numbers for Access, Volume, Checking, Initiation and Memory. Four minutes. Answer for this month, not for the version of your business you are planning. - Circle the lowest one.
That letter is your level, and it is also your next project. If two letters tie, pick the cheaper one to fix — in practice C and M are almost always the right pick, because checking and memory unlock everything above them. - Open the one card your letter points to.
Block 6 has the table. Use the prompt one level above where you are, not the most impressive one on the card. - Write one sentence and put a date on it.
“In 30 days, I will ______.” One move, not five. The other four letters are already ahead of your lowest and cannot deliver until it moves. - Re-run the Kitchen Check in ninety days, and tell me the two numbers.
Message me your lowest letter before and after. If it moved, I would like to know what moved it — that is the data this framework is short of for Malaysian SMEs.
Your trainer
Kalmen Chia — Co-Founder, E-Global SCM & DigitalAI Business Club. HRD Corp accredited trainer (Train the Trainer). Delivering ERP and AI automation for Malaysian SMEs since 2008, across Johor, Sabah, KL and Penang.
WhatsApp: +6016 773 7850
Fix One Letter. Not Five.
WSS-04 is free, and it is deliberately short. If your lowest letter turned out to be C or M — checking or memory — that is precisely what the hands-on Claude Mastery programmes are built to remove, and it is the wall most SMEs hit about thirty days after a session like this one.
Next sharing: REPLACE-INTAKE-DATE · Free member sharings run in the Webinar Short Sharing Series.
WhatsApp Kalmen: +6016 773 7850 See the 3-Level PathwayFrequently Asked Questions
Click any question to open it.
What is WSS-04, Where Are You on the AI Ladder?
WSS-04 is a free 90-minute online sharing for SME owners, directors and department managers. It walks the five levels of Claude adoption, from Level 0 The Advisor to Level 4 The Franchise Owner, then has every participant score their own business on the five-question Kitchen Check and leave with one 30-day move. Seven teaching blocks, five stated learning outcomes, and 12 prompt cards to take away.
Do I need to be technical to join this session?
No. The stated assumption is that you have used ChatGPT, Claude, Gemini or Copilot in a chat box, and nothing more. Nothing is installed, no account is required and no code is written. The five Kitchen Check questions are written without technical vocabulary on purpose, because the thing being measured is behaviour rather than tooling.
What is the Kitchen Check?
A five-question self-assessment scored 0 to 4 on Access, Volume, Checking, Initiation and Memory. It takes four minutes and needs no technical vocabulary. It is deliberately crude, and it is meant to be re-run every 90 days: a five-question check gets used quarterly, a fifty-question audit gets abandoned. It is a conversation instrument, not a certification.
Why is my level my lowest score instead of my average?
Because a chain does not hold at its average strength, and neither does a workflow. If AI drafts eight quotations this afternoon and you still read every line before any of them go out, the same number of quotations leaves the building as if it had drafted one. The parallel capacity produced nothing, because the checking step never scaled. The rule is also the kindest part of the framework: it means you have one thing to fix, not five.
What do I leave the session with?
Six things. Your five Kitchen Check scores and the lowest letter that names your level. The bottleneck holding you there, and the guardrail that removes it. One request from your own week, rebuilt using Goal, Context, Source, Constraints and Output. One 30-day next-level move, written as a single sentence. The 12-card prompt pack on this page. And the one-page My AI Context template. The deck, the Kitchen Check scorecard and the 30-day planner are sent to everyone afterwards.
Is this session HRD Corp claimable?
No. WSS-04 is a free, informal member sharing in the Webinar Short Sharing Series. It is not a registered HRD Corp programme, there is no fee, and no claim evidence is collected. The hands-on Claude Mastery programmes that follow it are registered separately. Claiming for those is done by the employer through their own HRD Corp application and remains subject to HRD Corp approval and the prevailing circulars.
Our customer data is confidential. Can we still use any of this?
Yes, and the session treats that as the right instinct rather than an obstacle. Start on non-sensitive data: price lists and product specifications before customer records, draft emails before signed contracts. Write the one-page policy on what AI may and may not touch first, because that document is the Level 0 to Level 1 guardrail. Follow your own company policy and your PDPA obligations. Every sample block on this page uses fictional data.
How is this different from the other 5 levels of AI posts going around?
Most of them repost the ladder. This session adds the instrument and the engine. The Kitchen Check turns the ladder into five observable behaviours you can score in four minutes, and the second half maps every bottleneck on the ladder to either a context problem or a checking problem, with the specific move that removes it. The ladder answers where am I. Context engineering answers how do I climb.