AI for Administrative Staff
Module 1
Illustration of an office worker chatting with an AI assistant

Getting Started with Generative AI

Before the tools, the mental model. Once you understand what these assistants are actually doing, good prompting stops being a trick and becomes common sense.

What generative AI is

A generative AI assistant is a very well-read writing partner that predicts the most suitable next words based on everything you have told it. It is excellent at language: rewording, structuring, summarising, translating, changing tone, and turning rough notes into a proper draft.

  • It is a language machine, not a records system. It writes well; it does not know your office.
  • It has no access to your files, your circular numbering, or last year's decisions unless you give them to it.
  • It produces the most plausible answer, not the verified answer, plausible and correct are not the same thing.
  • It does not remember you between separate chats unless the tool has a memory feature switched on.

What generative AI is not

Four things it cannot do for you

  • It is not a search engine with authority. It can state a policy clause that does not exist, in perfect officialese.
  • It is not a decision maker. Approval, exemption and disciplinary wording remain human decisions.
  • It is not a safe place for confidential data. Names, IC numbers, salaries and unpublished decisions do not belong in a chat box.
  • It is not accountable. If a wrong figure goes out in a circular, the signature at the bottom is yours.

Which assistant should I open?

All three do the same core job. The differences that matter to administrative work are small but real, and you only need one to start.

 ChatGPTClaudeGemini
Best atFast everyday drafting, rewriting, brainstorming, quick tablesLong documents, careful tone, natural official EnglishAnything already inside Google Docs, Sheets, Gmail and Drive
Office use caseEmails, memos, replies to enquiries, meeting agendasReviewing a 20-page report, polishing formal lettersDrafting inside a Doc, summarising a Gmail thread
Watch out forConfident wrong facts; over-enthusiastic toneSlower for one-line tasks; fewer built-in extrasOutput quality varies between Workspace surfaces
Free tierGenerous, enough for daily draftingUsable with daily limits on long chatsAvailable with a personal Google account

Choosing without overthinking

  • Pick one assistant as your default for a fortnight. Fluency with one beats shallow use of three.
  • If your office runs on Google Workspace, Gemini removes the most copy-and-paste.
  • If your work is long formal documents in English or Bahasa Malaysia, Claude usually needs the least editing.
  • Use a second assistant as a reviewer: draft in one, ask another to critique it.

The prompt formula: Role, Task, Context, Format

A weak prompt gets a generic answer. Four ingredients turn it into something you can almost send as-is.

  • Role: who the AI should be. "You are an administrative officer at a private university."
  • Task: the single, specific thing you want. "Draft a circular announcing a change of office hours."
  • Context: the details only you know: dates, department, audience, reason, tone, language.
  • Format: how the answer should look: length, bullet points, table, a subject line, a formal letter layout.
Weak prompt vs. strong prompt
WEAK:
Write an email about the office hours change.

STRONG:
You are an administrative officer at a private university in Malaysia.
Task: draft an internal email announcing new office counter hours.
Context: the counter will open 8.30 am - 4.30 pm from 1 September, closed 1.00-2.00 pm for lunch. Reason: staff rotation for a new service system. Audience: all faculty staff. Tone: polite, formal, brief.
Format: subject line, then 3 short paragraphs, then a closing line offering help at extension 1234.

Notice that nothing here is clever. It is simply the information you would have given a new colleague.

Context engineering: start long, then go short

The first message in a thread carries the heaviest load. Load it with role, task, context and format, then use short, precise follow-ups to steer the rest of the conversation.

  • A thread with full context at the start keeps the AI aligned without you repeating yourself.
  • After the opening, one-line follow-ups are usually enough: 'make it shorter', 'use British English', 'add a deadline column'.
  • This is context engineering: invest words once at the start, then spend them sparingly.
  • Save your best opening prompts as custom instructions or Project instructions so every new chat begins with the right context.
Thread that starts long, then goes short
You are an administrative officer at a private university in Malaysia. I need help drafting internal communications.
Task: turn my rough notes into a polite, formal email.
Context: audience is faculty staff, format is subject line plus 2-3 short paragraphs.
Rules: do not invent policy names, dates or approvals. If a detail is missing, write "[to confirm]".

Wait for my notes.


After this opening, your follow-ups can be tiny: 'more direct', 'add a deadline', 'Bahasa Malaysia version'.

Humanizer custom instruction
Write like a real person, not a chatbot. Use British English. Vary sentence length: mostly short, with the odd longer one for rhythm. Never pad.

Things to avoid completely:
- Closers that summarise what you just said or end with a tidy moral. Just stop when you are done.
- The "it's not just X, it's Y" construction.
- Repeated use of three-part lists used for effect. Two examples or four are fine.
- Em-dashes as a default connector. Use full stops or commas. Use em-dashes only when told or needed.

Things to do:
- Get to the point in the first line.
- Let sentences or paragraphs be uneven. No consistent repeated structure. Real writing is not perfectly balanced.
- Pick one good word over a string of adjectives.
- Skip the Oxford comma.
- Use a concrete example or a small analogy when it helps rather than abstract description. Be precise.

Paste this into ChatGPT Custom Instructions, or into a Claude Project instruction, and every draft will come back less robotic.

Thread set up once: public complaint replies
In this thread, I will be giving you complaint emails from members of the public and you must reply politely according to this flow:
- greet them
- state you take note or understand the situation
- offer possible solutions or indicate who to contact
- end with my details: Kee Man - Sarawak Skills People Happiness Officer, kmchuah@ss.edu.my, 012345678

Understood?

From then on, this thread becomes your reply station: paste each complaint email, get a ready-to-send reply in the same format every time.

Thread set up once: weekly summary reports
In this thread, I will paste my raw weekly activity notes and you will turn them into a summary report for management.
Rules:
- organise the notes under these headings: Key Activities, Issues Raised, Follow-ups Needed
- keep each bullet to one line
- do not invent activities or numbers that are not in my notes; write "[to confirm]" instead
- end with a one-sentence overall status: On Track, Needs Attention, or Urgent

Understood?

Set the thread up once at the start of the week, then every paste-and-send session takes under a minute.

Make your context do the work

  • Put the humanizer instruction at account level so every chat inherits it.
  • Start each new work thread with one big prompt that names the role, task, context and format.
  • Keep follow-ups short. If you find yourself rewriting the whole prompt, the opening probably missed context.
  • Update your custom instructions whenever you spot a phrase that still sounds like AI.

Copy-ready prompts for daily admin work

Official email from rough notes
You are an administrative officer. Turn my rough notes into a polished internal email.
Notes: [paste your rough points here]
Audience: [e.g. all academic staff / heads of department]
Tone: formal but warm, no jargon.
Format: subject line, greeting, 2-3 short paragraphs, clear action required with a deadline, closing with my contact details.
Keep it under 180 words.
Circular / memo drafting
Draft an official circular for internal distribution.
Subject matter: [what is being announced]
Effective date: [date]
Reason: [why this change is happening]
Who is affected: [group]
What they must do: [action, if any]
Format: reference line placeholder, title in capitals, numbered paragraphs, a closing paragraph with the contact person.
Language: [English / Bahasa Malaysia]. Keep sentences short and neutral.
Reply to a difficult enquiry
Help me reply to this message professionally without agreeing to anything I have not confirmed.
Their message: [paste]
My position: [what I can and cannot do]
Tone: calm, respectful, firm.
Format: 3 short paragraphs. Acknowledge their concern, explain the process, state the next step and the timeline.
Do not invent any policy, rule or approval that I have not stated above.

That last line is the single most useful sentence in this whole module. Reuse it everywhere.

Tone and language switch
Rewrite the text below three ways, keeping all facts identical:
1. More formal, for a letter to an external agency.
2. Simpler and friendlier, for an internal WhatsApp announcement.
3. In Bahasa Malaysia, formal register suitable for an official circular.
Text: [paste your draft]
Reusable template builder
Create a reusable template I can keep for future use, based on this finished document: [paste]
Replace all specific details with clearly labelled placeholders in square brackets.
Add a short note under each section explaining what belongs there.
Format: plain text I can paste into a Word document.

Using Projects in ChatGPT and Claude

Both ChatGPT and Claude let you create Projects. A Project is a dedicated workspace with its own instructions, files and chat history. It is the closest thing to giving the AI an office desk and a filing cabinet of its own.

  • Attach your most-used references: style guides, previous circulars, SOPs, templates, official signatures.
  • Set a standing instruction once: tone, format, language, and things the AI must never invent.
  • Keep all related chats in one place so you can find last month's draft without scrolling through a mixed inbox.
  • ChatGPT and Claude will apply the Project context to every new chat inside that Project.
Project instructions for an admin assistant
You are an administrative assistant in a Malaysian private university. Follow these rules for every task in this Project:

Tone: polite, formal, plain language, no jargon or AI-sounding phrases.
Language: use UK English unless I ask for Bahasa Malaysia.
Format: use our standard memo layout (reference line, title in capitals, numbered paragraphs, closing with contact person).
Rules:
- Never invent circular numbers, policy names, dates, names, statistics or approvals.
- If a detail is missing, write "[to confirm]" instead of guessing.
- Keep emails under 180 words and circulars under 300 words unless told otherwise.
- End official replies with my contact: [your name], [position], [email], [extension].

If you are unsure about anything, ask me before continuing.

Paste this into the Project instructions field once, and every chat in that Project will inherit it.

Project ideas for your unit

  • Create one Project for circulars and memos, another for event communications, and a third for replies to common enquiries.
  • Upload a few real (non-confidential) examples as reference files so the AI learns your house style from actual work.
  • Update the instructions whenever you spot a recurring mistake, this is how you train the AI cheaply.
  • Do not upload confidential documents. Use public templates, scrubbed samples, or your own wording only.

AI automation for repetitive admin tasks

Beyond one-off drafting, AI can act as a small production line for tasks that follow the same shape every week. You still set the rules, but the assistant does the repetitive language work.

  • Standard acknowledgement emails for common enquiries.
  • Formatting scattered notes into a clean table or report.
  • Turning a single announcement into versions for email, WhatsApp, and notice boards.
  • Generating monthly summary paragraphs from bullet-point updates.
  • Creating reusable templates from documents you have already written.
Batch email replies from a list
I need to send similar replies to several people. Do not invent any facts.

My standard reply template:
[ paste the reply you normally send ]

List of recipients and their specific details:
1. [Name] - [specific detail, e.g. requested document X]
2. [Name] - [specific detail]
3. [Name] - [specific detail]

Task: generate a separate, polite reply for each person. Personalise only the details I provided. Keep the tone and length similar to my template. Output each reply with the person's name as a heading.

This works for any repeated message: room bookings, leave approvals, document collection reminders.

Turn messy data into a clean table
I will paste messy notes below. Turn them into a clean, consistent table with these columns: [list columns, e.g. Date, Item, Owner, Status].

Rules:
- Use the same date format for every row.
- If a status is unclear, write "To confirm" rather than guessing.
- Highlight any rows that are missing required information.

Notes:
[paste your messy notes here]
One announcement, multiple channels
I need three versions of the same announcement.

Original announcement:
[paste the announcement]

Version 1: Formal email to senior staff, maximum 120 words.
Version 2: Friendly WhatsApp message for internal group, maximum 60 words.
Version 3: Short notice-board poster text, maximum 40 words, no emojis.

Keep all key facts identical across the three versions. Do not add information that is not in the original.

When to automate, and when to stop

  • Automate internal drafts and formatting, never external approvals or policy decisions.
  • Always keep a human checkpoint before anything leaves the office.
  • Save your best automation prompts as templates, so the next round is one paste away.
  • If the input changes often, automation saves time. If every case is unique, automation just adds steps.

Hands-on activity

  1. 1Open your chosen assistant and paste the 'Official email from rough notes' prompt.
  2. 2Fill it with a real (non-confidential) task sitting in your inbox right now.
  3. 3Read the output and fix three things: any invented detail, the tone, and anything too long.
  4. 4Ask a follow-up: 'Make paragraph 2 shorter and more direct.' Notice you never need to start over.
  5. 5Now run the 'Circular / memo drafting' prompt for a real announcement and compare the effort to doing it by hand.

You should end up with: One email and one circular drafted, plus a saved prompt you can reuse tomorrow.

Do

  • Give the assistant the same briefing you would give a new colleague.
  • Iterate in small follow-ups instead of rewriting the whole prompt.
  • Keep your best prompts in a personal notes file for reuse.
  • Read every line before it leaves your outbox.

Don't

  • Paste names, IC numbers, salaries or unpublished decisions into a chat.
  • Ask it to 'check the policy', it does not have your policy unless you attach it.
  • Accept the first draft because it sounds confident.
  • Let AI-flavoured phrases like 'delve into' and 'in today's fast-paced world' survive into official text.

Tips & pitfalls

  • If the answer is too generic, you left out context, usually the audience and the reason.
  • If the answer is too long, say the word limit out loud in the prompt. It obeys numbers well.
  • Ask 'what information do you still need from me?' before a complex drafting task.
  • Paste your office's previous circular as a style sample and say 'match this format exactly'.
  • End risky prompts with: 'Do not invent facts, names, dates or policies.'