Mister Mobile × Claude AI · Internal

The AI Playbook v2

v1 told you what was possible. v2 shows you what the team has already built — seven real cases, with the actual numbers, from people sitting in this office.

Version 2.0 · August 2026 Data: 3 Jul – 1 Aug 2026 Internal — HQ staff & HODs only
Why there is a v2

One month ago this was a plan. Now it's a list of things that exist.

v1 was written when almost nobody here was using AI, so it argued from what could work. That argument is over. In the last month an intern shipped a tool customers use to get repair quotes. An Ops colleague produced a 25-page training deck in two languages, on his own. Marketing built a cast of brand characters that now goes into every campaign. The ads budget is reviewed weekly against a dashboard that updates itself.

The one-line version

The same people, in the same week, are getting through work that used to take four to twenty times longer. The only thing separating them from everyone else is that they started.

What v1 got right

Chrome, Cowork and Chat are still the tools most people need. Start there.

What v1 missed

A fourth way of working — Claude Code — used by two of the seven cases below. You don't need it to start, but it's why the biggest results are as big as they are.

What hasn't changed

AI does the preparation, gathering, drafting and checking. You make the decisions and own the outcome. That won't change.

30 days · 3 July – 1 August 2026

What the team actually did last month

63
People with access
90%
Used it at least once
17,463
Messages exchanged
5,255
Cowork sessions run
5,069
Files created or edited
19
Median active days per person

Nineteen median active days out of a possible ~22 working days. For the people who have started, this is not an experiment they open occasionally. It is part of the working day.

Three different ways people work

One of the most useful things we learned this month: there isn't one way to use Claude. There are three, and they suit completely different people. Judging someone by the wrong one gets you the wrong answer.

💬 Chat 81%
🖥 Cowork 65%
⌨️ Claude Code 19%
Two or more 67%
Where we stand

Adoption snapshot

Good overall. One clear gap at the bottom.

41%
Heavy — using it most days
37%
Steady — using it most weeks
13%
Light — tried it, stopped
9.5%
Never opened it — 6 of 63

78% of the company uses this every week or more. That is a strong first month. The credit goes to the people in the cases below — their colleagues saw them doing it and copied them.

Six accounts have never been opened

Created, never used — not a single session in any tool. Another eight tried it in their first few days and never came back.

We are not listing names and nobody is in trouble. But we do want everyone past this point, because the gap between the top 41% and the bottom 22% is getting wider every month.

The usual reason people stop is that the first attempt wasn't faster than just doing the job by hand. That happens to nearly everyone — including all seven people in the cases below. The difference is they went back and tried again. It is the second and third attempt where it starts paying off. Pick the case closest to your job, copy what they did, and give it one more go this week.

HODs — this part is for you

Each HOD has been sent their own team's list privately.

If someone on your team is on it, give them fifteen minutes this week and walk them through whichever case below looks most like their job. That is usually all it takes. If they genuinely don't need the seat, tell us and we'll pass it to someone who does.

Not sure where to start with your team? Come and find Kai. Happy to sit in on a session, help pick a first use case, or set up access to whatever your department needs.

The evidence

Seven things built in the last month

Each of these was done by someone at Mister Mobile, in their normal job, alongside their normal work. Ordered from most approachable to most advanced — and the first one needs no technical skill at all.

These seven are examples, not a leaderboard

Plenty of other people have done excellent work with these tools this month — across Marketplace, Customer Service, Logistics, Finance, HR and more. We picked seven that show clearly different ways of working, so there is something here close to everyone's job. If yours isn't shown, that is not a judgement on it.

Send your case in for v3. What you did before, what you do now, and roughly what it saved. That is all we need.

A note on the time savings below: these are estimates, agreed as rough figures rather than measured with a stopwatch. They are meant to show the scale of the difference, not to be precise.

CASE 01
Kim
Operations · Bilingual training material
🖥 Cowork

What he made: a 25-page Trade-In System Upgrade training deck for Retail & CSS — covering the move from one device per reservation to up to ten, across WooCommerce, the Chat App and the Check-In App — in both Chinese and English, kept perfectly in sync.

Before

Screenshot every screen. Crop each one. Resize. Draw the red box on the right button by hand. Place it on the right slide. Then do the entire thing again for the second language, and again every time the system changes.

After

Upload a screenshot and describe what you want. It finds the right image, adds the red box, writes the explanation, and inserts it at the correct point — in both decks at once.

What was actually typed

“照片太小了。原本是 4 个在 1 个 page。你分成 4 page”

“The photos are too small. It was 4 on one page. Split it into 4 pages.” Ordinary words, in his own language.

Time saved ~2 days → ~2 hours estimated · 25 pages, twice, in two languages
"When I make a PPT, I just tell AI what I want… If photos are too small, I say 'split into 4 pages' and it's done. It even updates both Chinese and English decks at the same time. Before, I had to crop, resize, mark up, and place every image one by one. Now one sentence does all the steps. Faster, fewer mistakes, and I stay focused on the content." — Kim, Operations
What was typed One sentence — the photos are too small, split them across four pages.
Typed in Chinese, in the middle of the job.
What came out The finished, branded, version-numbered training deck.
The finished, branded, version-numbered training deck.
CASE 02
Jenny
Marketing · Content & brand assets
AI image generation

What was built: a reusable cast of Mister Mobile characters — the technician, the retail advisor, the delivery rider, the mascot, the auntie, the NS boy, the paranoid dad — that can be dropped into any storyboard and come out looking consistent every time.

Storyboard before video

Nine-panel cinematic storyboards are generated first, with camera and movement notes per panel. The storyline gets agreed and refined before anyone spends time or credits on video generation.

Product images without a shoot

Rather than photographing every SKU, suitable reference images are re-coloured and restyled to match our brand look, so they sit consistently alongside everything else on the store.

What was actually typed

“Replace the colour and outlook on the first image but keep the same placement like the second image”

One sentence — no studio, no camera, no editing software.

Time saved ~3 days → ~1 hour estimated · per storyboard set, vs briefing and illustrating it
"By creating the storyboard upfront, we can visualise the storyline more clearly, refine ideas earlier, and save both time and credits before moving into full video generation." — Jenny, Marketing
Storyboard first A nine-panel storyboard of an in-store scene, with the written instruction shown beside it
The instruction is on the right, the nine finished panels on the left. The story gets agreed here — before anyone spends time or money on video.
No photoshoot needed An iPad image recoloured to a different finish, with the one-line instruction shown beside it
The reference photo becomes an on-brand product shot. No camera, no studio.
CASE 03
Janet
Performance Marketing · Paid media
🖥 CoworkMeta & TikTok connector

What was built: a live Meta Ads budget dashboard that refreshes from Meta every time it's opened, compares the last seven days against the previous seven, and outputs a specific budget decision for every campaign.

One week's output

48campaigns reviewed
13flagged to scale up
8flagged to cut back
1,859website purchases tracked

Each recommendation comes with the reason attached — "cost per lead up 74% week-on-week and leads halved, pull budget back", or "spending 100% of its cap with strong return, room to scale". Not a dashboard you have to interpret. A list of decisions to approve or reject.

What was actually typed

“Give me a weekly data analysis on current meta ads setting, tell me which campaign need to increase/decrease budget base on the performance”

That was the whole instruction. It then asked which account, what “weekly” meant, and how to deliver the result.

Time saved ~6 hours → ~30 min estimated · weekly review across 48 campaigns
What was typed Claude asking which ad account, what time window and what output format, before running the analysis
It asks before it acts, instead of guessing what you meant.
What came out The live dashboard. Green = give more budget. Red = cut. Each with the reason.
The live dashboard. Green = give more budget. Red = cut. Each with the reason.
CASE 04
Brandon Lee
Intern · Shipped a live customer-facing product
⌨️ Claude Code

What he built: the Mister Mobile Repair Checker — "Get a Repair Quote in 60 Seconds". Customers pick their device, describe the symptom, get an instant price, and book. It syncs live against our WooCommerce catalogue, has an admin console for the ops team, and reports on how it is used.

What was actually typed

“comparing these two codebases, tell me if there is any drift, and if there is, which version is technically behind and in which aspects”

Plain English. No commands, no code.

Time saved ~2 months → ~2 weeks estimated · concept to a working, live product
"Using Claude Code, an Intern is able to build software products and digital files without touching the code; all with the use of plain English. Claude will clarify issues before making automatic decisions." — Brandon Lee, Intern
What he asked Brandon asking in plain English whether two versions of the code differ, and Claude working through the check
A question in ordinary words. It checked properly before answering, rather than guessing — and said so.
What customers see The Repair Checker landing page: Get a Repair Quote in 60 Seconds
The live product customers use — a price estimate in four steps, no technical knowledge needed.
CASE 05
Ethan
IT · Rapid prototyping
⌨️ Claude Code

What was built: the Mister Mobile Unified Inbox prototype — "Every marketplace. One queue." Shopee, Lazada, TikTok Shop, Carousell and Reebelo conversations answered from a single inbox. Concept to working prototype in days, against several weeks to a couple of months for a small dev team. Roughly a 5–10× speed-up.

Instead of asking for the whole thing at once, the work was broken into stages — research, design, build, testing, security check — with a human reviewing before it moved on. Same way you would run it with a team.

What was actually typed

“Build a single inbox that pulls in Shopee, Lazada, TikTok Shop, Carousell and Reebelo conversations”

The starting instruction. Everything after it was conversation and correction.

Time saved weeks → days his own estimate · roughly 5–10× faster
"I stay in overall control, guiding each stage… while the agents do the heavy lifting and I iterate." — Ethan, IT
The working prototype The eight-stage pipeline in the terminal, with the working Unified Inbox sign-in page above it
The finished prototype — “Every marketplace. One queue.” Built by describing what it needed to do, stage by stage, with a review before anything moved forward.
CASE 06
Zyler
ECSS Lead · Chat quality scored every day, not once a week
🖥 Cowork

What was built: an automated Chat Quality (CQ) review for the ECSS team. Every customer chat on the e-commerce platforms is pulled in, scored against our own written rules, and turned into a daily and weekly report — with the KPI dashboard updating itself.

Before

The Backliner read through the chats by hand, scored them one at a time and built the weekly CQ report. Twelve hours a week, every week — and by the time the numbers were ready, the conversations were already days old.

After

A bot exports the day's transcripts into a spreadsheet, one conversation per row. The AI scores each one, writes the coaching point with the evidence attached, and updates the dashboard. Zyler reviews the output instead of producing it.

The score isn't the AI's opinion. It marks against five things the team had already written down:

  • ECSS Knowledge Base — whether the SOP was followed
  • CQ Rules & Knowledge Log — what counts as a pass or a fail
  • CQ Scoring Rubric — how the score is worked out
  • ECSS Staff Roster — who was actually on shift
  • Approved Chat Templates — whether the mandatory templates were used
Time saved ~12 hrs/week → ~1 hr/week estimated · and the review is now daily instead of weekly
"The weekly CQ report used to take the Backliner twelve hours. Now the scoring is done every day, against the same rules every time, with the evidence already attached. My time goes into the coaching instead of building the report." — Zyler, ECSS Lead
What the AI works from The CQ scoring sheet: one chat per row, with a score, a RAG flag, the issue type and a pass or fail against each of the nine rules
One chat per row, scored against every rule, with the issue type and the coaching note beside it. Green, amber or red at a glance.
What came out The ECSS customer service KPI dashboard showing first response time, resolution time, chat quality score and escalation rate per platform
The ECSS KPI dashboard — first response, resolution, chat quality and escalations, per platform, updated without anyone assembling it.
CASE 07
Kai
IT · The layer everything else runs on
Built by IT

What was built: a safe connection between Claude and our own systems — appointments, stock, products, orders, customers and website content — plus a control panel deciding exactly who can reach what.

This is why the other cases are possible — and why they are safe. Everyone is given access to exactly the systems their job needs, and nothing else. Nothing is switched on until someone decides it should be.

The access panel Access control panel listing each system separately, with the permissions that only read shown above the ones that change data
Each system listed separately, with the things that only look above the things that change. Nothing is ticked until someone decides it should be.
The method

What to copy from them

Seven different jobs, a handful of different tools. Underneath, all seven people worked the same way — and it comes down to three habits you can use tomorrow.

  1. Say what you want, not how to do it

    Kim never explained how to crop an image or where to put it. He said "the photos are too small, split it into 4 pages." Janet didn't specify a calculation. She said "tell me which campaign needs more budget." Describe the result you want in a sentence, the way you'd brief a colleague, and let it work out the steps. This is the single biggest difference between people who get value from it and people who don't.

  2. Expect the first answer to be wrong — and say so

    Ask for a job description and the first draft will read like every other job ad. Tell it so: "too generic — this role is 80% walk-in customers, put that first." Ask it to summarise last month's spend and it may group things in a way that doesn't match how you report. Say that, and it regroups. Nobody on this page got what they wanted on the first attempt. They pushed back in ordinary words and it corrected. If you stop at the first answer, you are getting a fraction of what's there.

  3. Build the thing that repeats, not the one-off

    Kim's deck rebuilds itself, so the English and Chinese versions stay in step for free. Janet's dashboard refreshes on opening instead of being rebuilt every Monday. Jenny's characters look identical across every campaign because the cast was saved once. Twenty extra minutes making something reusable is what turns a good afternoon into a permanent saving.

None of them started with a big idea

Every case here began as one annoying task somebody wanted off their desk. The training deck, the weekly budget review, the repair quote — all ordinary work that was taking too long. Pick the thing you have been putting off and start there.

The honest bit

Don't be afraid to use it

The most common reason people haven't started isn't that they tried and it didn't help. It's that they haven't tried. Usually one of these six worries is behind it — so let's take them one at a time.

"I'll look like I can't do my own job"

The seven people above are the opposite of that. An intern shipped a product. An Ops colleague produces bilingual training material single-handed. Using it well is a visible skill — that's precisely why they're in this document.

"I'm not technical enough"

Kim's case involves no code, no setup and no connectors. He uploads a screenshot and describes what he wants in a sentence, in whichever language he prefers. If you can brief a colleague, you can do this.

"What if it's wrong?"

Sometimes it will be — which is why you check it, exactly as you'd check a junior's work. Note that in Janet's case it flagged a flaw in our own data without being asked. Used properly it catches more mistakes than it makes.

"I don't have time to learn it"

Nobody on this page took a course. They had a task in front of them and tried describing it instead of doing it. The first attempt takes longer than doing it yourself. The fourth doesn't.

"Is it safe with our data?"

This is a fair question and it has a real answer — see the governance section below. Access is granted per person per tool, with daily caps. The rules exist so the answer can be yes.

"My work isn't on the list"

Then you've found the point. The department lists in v1 were a starting guess, not a boundary. It writes, checks, compares, summarises, reformats, translates, plans and builds. If a task is repetitive, or you've been putting it off because it's tedious — try it there.

The line that doesn't move

It expands what's possible. You still decide.

This matters enough to state plainly, because it's the thing that makes everything above acceptable.

What AI does

  • Gathers, compares and organises information
  • Drafts a first version of almost anything
  • Checks every item instead of a sample
  • Surfaces options you hadn't considered
  • Handles the tedious middle of the work

What stays with you

  • The decision — approving the budget change, sending the reply, publishing the price
  • The judgement — whether the output is actually right for this customer, this branch, this situation
  • The accountability — it is your work when it goes out, exactly as if you had typed every word

Put simply

"The AI wrote it" is not an explanation for a mistake, in the same way "my calculator did it" wouldn't be. Check the work before it carries your name. That standard is unchanged from before AI — the only difference is how much work you can now get through.

Governance — summary

The rules, in short

Full policy is maintained separately. This is what you need to know to work day to day.

  1. Claude is our default, approved platform

    Not because other tools are bad, but because it's the one we have an agreement with, control access to, and can audit. Approved means data-safe.

  2. Another tool is genuinely better? There's a path for that

    Preferring a different AI isn't on its own a reason. But where a tool does a specific job materially better — image generation is the obvious example — request it and it goes on the approved list. Two routes: either the tool itself is data-safe, or the process guarantees no sensitive data ever reaches it. Data safety is the floor in both cases, never traded away.

  3. Never enter these into any AI tool — including Claude

    Customer personal data · IMEI-linked personal records · payment or card details · staff records. This applies regardless of which tool you're using. The bigger risk isn't which AI you pick — it's what you paste into it.

  4. Personal accounts are the actual danger

    Using your own ChatGPT or Gemini login for company work means our data leaves for somewhere with no agreement, no controls and no record. That's a PDPA exposure. If you need something we don't have, ask — don't route around it.

  5. Access is per person, per tool, with daily caps

    Nothing is enabled by default. Ask IT for what your role needs.

Your move

Start here this week

If you've never opened it

Take the most tedious thing on your desk this week — the one you keep postponing. Describe it in one paragraph, the way you'd explain it to a new colleague. That's the whole first step.

If you tried it once and stopped

Find the case above closest to your job and copy the approach rather than starting fresh. Then go and ask that person — they're named here for exactly that reason.

If you're already using it daily

Turn one repeated task into something reusable — a saved project, a script, a dashboard. That's the step between the steady 37% and the heavy 41%.

For HODs

Three things this month:

  • Check whether anyone on your team is in the six never-activated seats, and sit with them for fifteen minutes.
  • Count all three ways of working before you judge someone's usage — chat, Cowork and Claude Code look completely different in the data.
  • Send one case from your department for v3. The cases here landed because they're real and local, not because they're impressive.