Kopi & Prompt ☕🌺

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 If your "to-review" pile is now tall enough to need its own parking space, this one's for you. This week we hand the first read of a student thesis, paper, or slide deck to NotebookLM — Google's tool that answers only from the document you upload and cites the exact page for every point. Think of it as a tireless teaching assistant who has actually read the whole thing, never sulks, and never once asks for a teh tarik break. 🤖☕ You stay the expert; it just clears the runway.

Built on the RACE frame (Role · Action · Context · Expectation) + a critical-thinking checklist. Copy it, fill the [BRACKETS], paste into NotebookLM.

💡 Follow-ups that work well: "Expand point 3(c) with examples," "Turn the priority list into a student checklist," or "Draft two encouraging sentences for the cover note." Alternatives: build a reusable Gem in Gemini with this prompt as its instructions; use Gemini in Docs for inline comments once you trust the pass.

Teachers who use AI report real gains in time and quality — the reason a fast, careful first read pays off.

Source: Gallup & Walton Family Foundation, Teaching for Tomorrow: Unlocking Six Weeks a Year With AI (2025); n=2,232 US K-12 teachers, ±2.5pp. (K-12 sample — closest large-scale evidence; indicative for higher ed.)

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 We all know the sacred ritual: a new program is approved, and you open last year's kertas kerja, "Save As", and begin the archaeological dig — hunting for every place the old dates, old budget, and (horror) the previous program's name are hiding. 😅 This week we retire the copy-paste treasure hunt. Hand Gemini last year's paperwork as a template, give it a few bullet points about the new event, and it rebuilds the whole document in the same house format — sections, tone, tables and all. You still own the numbers and the sign-off; Gemini just does the tedious re-typing.

Technique this week: template-transfer (document-grounded) prompting — you give the AI a real past document as the gold template, plus a few facts, and it rebuilds it. Attach last year's kertas kerja, then paste this.

💡 Handy follow-ups: "Draft the surat kelulusan cover note," "Build a atur cara tentatif (run-of-show) table," "Make a Jadual Anggaran Perbelanjaan from these items," or "Give me an English + BM version of the headings." Level-up: save the whole thing as a Gem called "Kertas Kerja Generator" so every future event starts from one click. Strict grounding? Use NotebookLM instead so it only pulls from the old document.

Paperwork is a real slice of academic time — the case for making the first draft fast.

Source: Federal Demonstration Partnership 2012 Faculty Workload Survey (US National Academies); n≈13,000 principal investigators. (US research-admin sample — indicative of the paperwork load faculty carry.)

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 It's assessment season, and the blank question paper is staring back at you — again. Writing a dozen fair, well-aligned questions from scratch (with plausible distractors, no less) is the kind of task that quietly swallows an evening and a pot of kopi. 😅 This week we hand Gemini your lecture notes and let it draft a whole question bank — mapped to your learning outcomes and Bloom's levels, complete with an answer key. You stay the examiner-in-chief; Gemini just fills the first draft so you can spend your time sharpening, not staring.

Technique this week: few-shot + specification prompting — show one model question in the exact format you want, then specify the mix (how many, which Bloom's levels, which outcomes). Paste this into Gemini with your notes.

💡 Handy follow-ups: "Convert these into a Google Forms quiz with auto-marking," "Add 3 harder Analisis-level items," "Write a marking scheme for the short-answer questions," or "Give me a fully Bahasa Melayu version." Level-up: save it as a Gem called "Bank Soalan" and grow a reusable question pool each semester. Tip: feed it the topic AND your CLOs together — alignment gets much sharper.

In a blinded multinational trial, AI drafted exam questions in a fraction of the time — at quality assessors couldn't reliably tell apart.

Source: Cheung et al., "ChatGPT versus human in generating medical graduate exam MCQs," PLOS ONE (2023); multinational (HK/Singapore/Ireland/UK), 100 MCQs rated by 5 blinded assessors. Note: a separate study found only ~32% usable without edits — human review stays essential.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Notice something about the last few issues? Each time, we typed out a careful prompt… and then next week, typed it out again. 😅 This week we break the loop. A Gem is a custom AI assistant you set up once — you save the instructions, give it a name, and from then on it's a single click in your Gemini sidebar. Better still, you can share it with the whole team, so everyone drafts the same way without re-explaining anything. This is the shift from clever prompting to renewed workflow: less re-typing, more actual work.

This week the "prompt" becomes a saved instruction set — you paste it into the Gem builder once. Here's a worked example: a Gem that replies to student emails.

💡 Turn the whole series into Gems: paste any earlier prompt — Kertas Kerja, Bank Soalan, minit mesyuarat — as a Gem's instructions and it becomes a one-click tool. Add rules over time like "always sign off with my name" or "reply in BM only." Note: sharing a Gem across FTSM may need your IT admin to enable Gem sharing in Workspace — worth a quick check before you build the team version.

So much of the week goes to repeating and re-coordinating work — the very drudgery a reusable Gem removes.

Source: Asana Anatomy of Work Index 2021; 13,000+ global knowledge workers. (Cross-industry sample — the duplication it measures maps closely to academic re-work: the same feedback, paperwork, and prompts, re-written each time.)

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The email arrives: "Revise and resubmit." You open the file and there it is — three reviewers, forty-one comments, and Reviewer 2 in a mood. 😅 The science is the easy part; the hard part is replying to every point clearly, tracking each change, and staying gracious when you'd rather argue. This week we let Gemini turn that wall of comments into a tidy point-by-point response table — conceding where the reviewer's right, pushing back politely where they're not. You keep the scientific judgment; Gemini keeps you organised and diplomatic.

Technique this week: persona + tone-constraint prompting — you set a role (a measured corresponding author) and hard tone rules (concede graciously, disagree politely, never dismissive), then force a complete table. Paste it into Gemini with the reviews.

💡 Handy follow-ups: "Draft a short cover letter to the editor summarising the main changes," "Make my response to comment 3 firmer, but still courteous," "Turn the table into a formatted response letter," or "Give me a Bahasa Melayu version." Level-up: save it as a Gem called "Response to Reviewers" (see Isu #04) so every resubmission starts the same way.

Publishing runs on the revise-and-resubmit cycle — slow and multi-round. A clean reply that settles comments the first time can save a whole round.

Source: Publons Global State of Peer Review (2018); 11,800+ researchers — median 16.4 days per review round, ~5 hours to complete a review. Cumulative days shown at the median rate; most manuscripts undergo multiple rounds before acceptance.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Your paper is published — congratulations! Now the editor of a magazine (or your university's penerbitan popular) wants "a short piece the public can actually understand." Cue the panic: how do you explain three years of work to someone who has never heard of your field, without your abstract's forty-syllable words? 😅 This week we let Gemini translate your research into plain language — a proper article with a hook, a headline, and zero jargon — while keeping every finding accurate. You supply the science; Gemini supplies the readability.

Technique this week: audience-transformation prompting — you name the reader and the reading level, and tell the model what to preserve while it changes the register. Paste it into Gemini with your abstract.

💡 Handy follow-ups: "Give me 3 punchier headline options," "Rewrite it fully in Bahasa Melayu for a local magazine," "Add a 100-word version for social media," or "Suggest one everyday analogy to explain [concept]." Bonus: the same rewrite doubles as a grant impact statement — ask for "a 150-word version for a funding application." Level-up: save it as a Gem called "Lay Summary" (see Isu #04).

Academic writing keeps getting harder to read — which is exactly why a plain-language version earns its keep.

Source: Plavén-Sigray et al., "The readability of scientific texts is decreasing over time," eLife (2017); 709,577 abstracts, 1881–2015. Flesch Reading Ease (higher = easier): 60–70 = plain English (age ~13–15); 30–50 = college; 0–30 = graduate.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Every semester, the same plea lands in your inbox the week before the exam: "Prof, boleh bagi study guide tak?" 😅 Building one by hand — key concepts, definitions, practice questions — is a genuine evening's work. This week we let NotebookLM do it: upload your readings and lecture slides, and with a click it produces a study guide, an FAQ, and even an Audio Overview — a two-host "podcast" your students can listen to on the bus. Best of all, the guide bakes in self-testing questions, which is exactly what the learning science says makes revision stick.

Technique this week: app-feature prompting — you drive NotebookLM's built-in outputs (Study Guide, Audio Overview) and steer them with a short instruction. Upload your sources first, then use these two.

💡 Handy follow-ups: "Turn the study guide into a printable one-pager," "Generate 15 flashcards (term → definition)," "Make a timeline of the key ideas," or "Create a Bahasa Melayu version of the guide." On the audio: you can now generate the Audio Overview in Bahasa Melayu too — worth trying for local cohorts. Note: students don't need a NotebookLM account to open a shared, view-only notebook link.

Why the guide includes self-tests: recalling beats rereading. That's the pedagogy behind the practice questions.

Source: Roediger & Karpicke, "Test-Enhanced Learning," Psychological Science (2006). Note: right after studying, rereading looks better — but the advantage flips with delay, which is what matters for exams.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Every literature review begins the same way: forty PDFs open in your browser, a blank table, and the slow, soul-sapping copy-paste of "who did what, found what, missed what." 😅 By paper number twelve you've forgotten what paper number three even argued. This week we let NotebookLM build the synthesis matrix for you — one row per paper, columns for aim, method, findings and gaps — then surface the themes, the disagreements, and the hole in the field. Because it's grounded in the papers you upload, every cell is cited back to its source. You judge the quality; NotebookLM does the wrangling.

Technique this week: comparison-matrix prompting — you define the columns once and let the model extract the same fields from every source. Upload your papers to NotebookLM, then paste this.

💡 Handy follow-ups: "Sort the matrix by method," "Add a column for theoretical framework," "Draft a 200-word synthesis paragraph from the 'agreement' themes," or "Flag any two papers with directly conflicting findings." Few papers only? Gemini can do the same from attached PDFs — but for many sources, NotebookLM's grounding keeps citations trustworthy. Level-up: save the prompt as a Gem called "Synthesis Matrix" (see Isu #04).

No one can read it all — the literature grows exponentially, which is why a matrix that imposes order is worth the effort.

Source: Bornmann, Haunschild & Mutz, "Growth rates of modern science," Humanities & Social Sciences Communications (2021); overall growth 4.10%/year, doubling time 17.3 years. Post-war output has at times grown ~8–9%/year — doubling roughly every 9 years.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 A long email thread lands about a program: dates half-agreed, three people tagging you, and a deadline buried somewhere in message eleven. 😅 You know that if you don't act now, it'll resurface as a panic next week. This week we let Gemini work across your Workspace: summarise the thread right inside Gmail, pull out the decisions and action items, then — from the Gemini app — spin up a Calendar event and a Task without you typing them twice. Less app-hopping, fewer dropped balls. You confirm; Gemini connects the dots.

Technique this week: cross-app prompting — Gemini reads one app and acts in another. First summarise inside Gmail, then use @app mentions in the Gemini app to create the event and task.

💡 Handy follow-ups: "Draft a reply agreeing to option B," "Find all unread emails from the Dean this week," "Create a task for each action item," or "What's on my calendar tomorrow?" Note: the @app actions rely on Workspace apps/extensions being switched on for your account — if you don't see them, ask your IT admin. Always let Gemini show the draft event or reply before it sends or creates anything.

Email and coordination swallow most of the week — the case for turning inbox time straight into action.

Source: McKinsey Global Institute, "The Social Economy: Unlocking Value and Productivity Through Social Technologies" (2012); IDC + MGI analysis of interaction-worker time use.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 We all know that a good case study turns a sleepy tutorial into a proper debate — but writing one from scratch (a believable scenario, some data to chew on, questions that build) is the sort of task you keep meaning to do and never quite finish. 😅 This week we let Gemini generate a full problem-based learning case: a real-world dilemma with a named protagonist, a small dataset of clues, group roles, and tiered discussion questions — plus a facilitator's note flagging the misconceptions to watch for. You bring the subject expertise and the facilitation; Gemini builds the scaffolding.

Technique this week: scenario-generation prompting — you specify the parts to generate (scenario, data, roles, tiered questions) and the constraints, and Gemini builds a whole teaching artifact. Fill the [brackets].

💡 Handy follow-ups: "Add a plot twist for the second half," "Generate a marking rubric for the group output," "Turn the dataset into a Google Sheet the groups can edit," or "Give me a fully Bahasa Melayu version." Level-up: save it as a Gem called "Case Study Builder" (see Isu #04) and spin up a fresh case for any topic in seconds.

Why bother with cases: active learning consistently beats straight lecturing — this is the evidence behind it.

Source: Freeman et al., "Active learning increases student performance in science, engineering, and mathematics," PNAS (2014); meta-analysis of 225 STEM studies. Exam scores also rose ~0.47 SD — about half a letter grade.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The grant deadline is in ten days, you have a page of brilliant bullet points, and a "Background & Significance" section that stubbornly refuses to write itself. 😅 We've all stared at that blinking cursor. This week's trick isn't to ask Gemini to "write my proposal" — it's to make it step back first: articulate the big-picture problem, the gap, and why your approach fits — then draft the section on that foundation. The result reads like an argument, not a list. You bring the ideas and the citations; Gemini builds the scaffolding and the prose.

Technique this week: step-back prompting — before drafting, you make the model establish the big-picture reasoning, then write on that foundation. Two moves, pasted one after the other into Gemini.

💡 Handy follow-ups: "Tighten this to 350 words," "Add a paragraph on Malaysian/regional relevance," "Turn my aims into 3 crisp objectives," or "Rewrite the significance for a non-specialist reviewer." Why step back works: forcing the big-picture logic first gives the draft a spine — importance → gap → solution — instead of a pile of facts. Level-up: save it as a Gem called "Grant Draft" (see Isu #04).

Grant writing is a huge time sink, and most of it isn't funded — so a fast, strong first draft genuinely pays off.

Source: Herbert et al., "On the time spent preparing grant proposals," BMJ Open (2013); NHMRC Australia, 285 researchers / 632 proposals — avg 34 working days each, 21% funded (~550 researcher-years in one round). Per-funded figure shown at the ~21% success rate.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Every OBE audit season, the same document haunts us: the Jadual Spesifikasi Ujian — the blueprint that proves your exam actually spreads its marks across your CLOs and Bloom's levels, and isn't just twenty recall questions in a trench coat. 😅 Building the grid by hand, with all the totals adding up, is fiddly arithmetic. This week we let Gemini draft the JSU: a proper CLO × Bloom's matrix with mark weightings that balance, higher-order coverage built in, and a version you can paste straight into Google Sheets. You set the weightings; Gemini does the grid and the sums.

Technique this week: structured-blueprint prompting — you fix the two axes (CLO × Bloom's) and the numeric constraints (weightings, totals), and Gemini fills and balances the grid. Paste this into Gemini.

💡 Handy follow-ups: "Add a column for item type (MCQ / structured / essay)," "Rebalance to 50% higher-order," "Generate a blank item-writing template for each cell," or "Give me a Bahasa Melayu version of the headers." Pairs with Isu #03: once the blueprint is set, use the question-bank prompt to draft the items that fill each cell. Level-up: save it as a Gem called "JSU Builder" (see Isu #04).

Left to habit, exams skew hard toward recall — which is exactly what a blueprint is designed to prevent.

Source: Momsen et al., CBE—Life Sciences Education (2010); 9,713 exam items from 50 undergraduate-biology instructors, classified by Bloom's level. A blueprint is how you deliberately shift that balance.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The first FYP meeting, every year: an eager student, a spark of an idea — "Prof, I want to do something with AI" — and a scope roughly the size of a PhD. 😅 Half of supervision is gently steering that enthusiasm into a project that's actually finishable in two semesters. This week we use Gemini as a scoping partner: first diverge — generate a spread of candidate topics from the student's interests and your research area — then converge — score each on scope, feasibility, data and novelty, and sharpen the best into a real problem statement. You keep the academic judgment; Gemini structures the conversation.

Technique this week: divergent–convergent prompting — first ask for many options, then filter them against explicit criteria. Two moves, pasted one after the other into Gemini.

💡 Handy follow-ups: "Add a risk & mitigation note for the top topic," "Suggest a public dataset or API for topic 2," "Draft a one-page proposal outline," or "Give me a Bahasa Melayu version." Why diverge then converge: generating widely first, then filtering against your criteria, beats settling on the first idea — and surfaces options the student hadn't considered. Coming in Isu #16: a supervision tracker to manage the project once the topic is set.

Most software projects don't fully succeed — and unclear scope is a leading reason. A tight FYP topic is half the battle.

Source: Standish Group CHAOS report (2015 database, FY2011–2015); incomplete, changing or unclear requirements are among the top causes of the challenged/failed share. Small projects succeed ~90% of the time vs under 10% for large ones.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Some tasks aren't hard — they're just relentless. Every Monday: what's due this week, which papers dropped in my field, what did I promise someone? 😅 Doing it by hand is five minutes you keep forgetting to spend. This week we go one step past the Gem from Isu #04: we schedule it. Save the routine as a Gem, then set Gemini to run it every Monday at 8am and hand you the digest — no clicking required. It's the difference between owning a tool and having the tool show up for work on its own. You keep the judgment; the routine keeps the time.

Technique this week: automation-scheduling — not a cleverer prompt, but a standing one that runs on a timer. Save the routine as a Gem, then schedule it. Two moves.

💡 Other routines worth scheduling: a Friday wrap-up of what shipped this week, a monthly "citations of my papers" check, or a daily 7am summary of your unread priority mail. Honesty note: Gemini's Scheduled actions is still rolling out across Workspace — if you don't see the clock icon yet, run the "Monday Digest" Gem manually each week and set a recurring Calendar reminder (same result, one tap). Ask your IT admin if scheduling isn't enabled for FTSM accounts.

Plenty of your week is routine that a machine could handle — but your job isn't going anywhere. Automate the tasks, not yourself.

Source: McKinsey Global Institute, "A Future That Works: Automation, Employment, and Productivity" (2017); analysis of 2,000+ activities across 800+ occupations. The most automatable category is data collection & processing — exactly what a scheduled digest does for you.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The work is done, the slides are ready — and then comes the part that keeps everyone up at night: the questions. That one examiner who asks, ever so gently, "but why did you choose this method?" 😅 The best cure isn't more re-reading — it's rehearsal under fire. This week we turn Gemini into a tough-but-fair examiner: it asks one probing question at a time, escalates, and gives feedback each round — so the real viva feels like the second time, not the first. Brilliant for prepping your students, and quietly handy for your own conference Q&A too.

Technique this week: role-play / simulation prompting — you cast the model in a role and set the rules of the exchange (one question at a time, stay in character, escalate). Paste this into Gemini.

💡 Handy follow-ups: "Now play a hostile reviewer," "Focus only on my methodology," "Ask the questions in Bahasa Melayu," or "List the 5 questions I answered worst, with model answers." Not just for students: run it on your own paper before a conference to rehearse the Q&A. Level-up: save it as a Gem called "Mock Viva" (see Isu #04) so any student can spin up a practice session.

The viva or defence is this fear at maximum stakes — which is why rehearsal, not just re-reading, matters.

Sources: Dwyer & Davidson (2012), college-student survey — ~61% cite public speaking as their most common fear; Marinho et al. (2017) — ~63% of the general population report some public-speaking anxiety (via a 2024 peer-reviewed review). A Spanish university study found >75% of students affected.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The supervision meeting ends, the student leaves, and you mean to type up the notes — but the next class starts, and by evening "she'll fix the methodology" has blurred into a vague good feeling. 😅 We built the tracker last time; this week we feed it by voice. Right after a meeting, mumble a 60-second note into your phone — and Gemini turns it into a tidy tracker row: decisions, action items, next deadline, at-risk flag. Talking is far faster than typing, and nothing slips through the cracks. (Want to capture the whole meeting? You can — with the student's consent. More on that below.)

Technique this week: voice-to-structured-log prompting — you hand Gemini a spoken note plus a fixed output schema, and it returns one clean row. New trick for us: the input is audio, not typing.

💡 Handy follow-ups: "Turn these 5 memos into 5 rows at once," "Draft the student's action-item reminder email," "Summarise this student's last 3 meetings," or "Give me the row in Bahasa Melayu." Don't have the tracker yet? The two-table supervision Sheet this fills lives in the parked tracker edition — ask Gemini to "build an FYP supervision tracker" first. Completes the FYP pair: Isu #13 scoped the topic; this one keeps it on track to submission.

Research projects stall for want of structure as much as ability — regular, recorded supervision is among the strongest predictors of finishing.

Source: Council of Graduate Schools PhD Completion Project (doctoral data; ~57% complete in 10 years). Supervisor-relationship difficulties are cited by ~25–30% of non-completers, and frequent (monthly+) supervision predicts completion (Pyhältö et al., 2015). The principle — regular, recorded supervision aids completion — applies to FYP too.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 AI writes you a beautiful paragraph, complete with a citation that looks perfect — right journal, plausible authors, tidy year. There’s just one problem: the paper doesn’t exist. 😅 These “hallucinated” references are the fastest way to lose a reviewer’s trust. This week we turn the tables with chain-of-verification: have the AI (or a second tool) check each citation against reality. Gemini with grounded search can look up whether a source is real; NotebookLM, working only from PDFs you actually uploaded, simply can’t invent one. Trust, then verify — every single reference.

Technique this week: chain-of-verification — instead of trusting the first answer, you make the model re-check its own claims one by one against a real source. Paste your reference list (or draft) into Gemini.

💡 Handy follow-ups: "List only the references you could not verify," "Give me the working DOI link for each," or "Rewrite this paragraph using only the three sources I uploaded." The golden rule: a citation an AI produced is a lead to check, never a source to trust — open the DOI and confirm it yourself. Belongs to the rigour thread: the same "trust, then verify" habit runs through Isu #20 (statistics) and #25 (reference consistency).

Even the best models invent references that look flawless — and a good share of the real ones still carry errors.

Walters & Wilder (2023, Scientific Reports 13:14045): GPT-3.5 fabricated 55% of citations, GPT-4 18%. Linardon et al. (2025): GPT-4o fabricated ~20%, and ~45% of genuine citations still had errors. Model and year matter — treat as a range.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Your student’s poster has twelve paragraphs, six tables, and a font size best viewed through a microscope. 😅 At a real poster session, though, people drift past and give each poster about a minute and a half — so a wall of text is a wall they’ll walk past. This week we help your student (or you) design a poster that leads with one big finding: use Gemini in Google Slides with a clear design brief — key message first, minimal text, a strong visual, clean hierarchy. Let the science shine; don’t bury it.

Technique this week: design-brief prompting — you don’t ask for “a poster,” you specify the one message, the audience, and a visual hierarchy, so the layout serves the finding. Use Gemini in Google Slides.

💡 Handy follow-ups: "Cut every section to 30 words," "Suggest a colour-blind-safe palette," "Draft a 30-second spoken pitch to go with it," or "Make a Bahasa Melayu version." Reality check: stand back two metres — if the finding isn’t readable from there, it’s still too busy. Pairs with the student arc: Isu #21 (mentoring) and #23 (reference letters) round out supporting your students.

Poster sessions reward one clear message, not a wall of text — you have seconds to earn a viewer’s attention.

Science Communication Toolkit / Principles of Scientific Communication (2020): viewers spend on average ~90 seconds per poster, and a poster should be understandable in under 5 minutes by a non-specialist. An AIChE poster analysis noted many decide in under 5 seconds whether to stop. Design guidance, not a controlled trial.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 A systematic review begins with a search that returns 1,200 hits, of which maybe 40 are relevant. 😅 Reading every title and abstract by hand is where reviews go to die (usually around abstract number 300, at 1am). This week we speed it up honestly: turn your inclusion / exclusion criteria into a rubric and have the AI apply it to each abstract as a first pass — including a “not sure” flag — while you make every real decision. A tireless assistant, not a replacement for your judgment.

Technique this week: rubric-as-classifier — you hand the model your eligibility rules and it labels each record against them, flagging the uncertain ones for you. Works in Gemini, or NotebookLM grounded to your export.

💡 Handy follow-ups: "List only the Unsure ones with the missing criterion," "Summarise why each Exclude failed," or "Export this as a table for my PRISMA log." Non-negotiable: AI screening is a second pair of eyes, never the decision-maker — keep a full audit trail and, for a real review, dual human screening still applies. Rigour thread: shares the "AI drafts, human confirms" spine with Isu #17 and #20.

Title-and-abstract screening is the biggest time sink in a review — and where AI as a careful first pass helps most.

Cochrane (2024 Global Evidence Summit): dual title/abstract screening averages ~114.6 hours (14.3 person-days) per review at ~36 sec/abstract. A 2024 LLM-assisted study (Rayyan) reported ~62% less screening time at one threshold. AI assists; humans still decide.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Somewhere out there, a Cronbach’s alpha is being calculated on four nominal categories, and a t-test is being run on data that has never once been normal. 😅 We’ve all been tempted to reach for the familiar test rather than the right one. This week we build a statistical red-team: describe your variables, design and planned analysis, and have Gemini check them against the rules — data level vs test, assumptions, the right reliability measure — and flag the mismatches. A sceptical second reader, before a reviewer becomes the first.

Technique this week: assumption-audit prompting — cast the model as a critical statistical reviewer with an explicit checklist, so it audits your design instead of just agreeing with it.

💡 Handy follow-ups: "Suggest the correct test and why," "What assumption checks should I run first?" or "Explain, in plain terms, why alpha is wrong here." Say it loudly: LLMs are not statisticians and can be confidently wrong — use this to catch obvious mismatches, then verify. Rigour thread: the auditing habit continues from Isu #17 (citations) into #25 (references).

Statistical slips are common even in published work — a pre-submission check catches the ones that change your conclusions.

Nuijten et al. (2016, Behavior Research Methods 48:1205–1226): across 250,000+ p-values in ~30,000 psychology papers, ~half had at least one inconsistency between a test statistic and its p-value, and ~1 in 8 had a gross inconsistency. International benchmark.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 “Prof, should I do a Master’s, find a job, or… both?” — usually asked at 5pm on a Friday, with genuine anxiety behind it. 😅 Good mentoring isn’t handing over an answer; it’s helping the student think clearly about their own options. This week we use Gemini as a decision-support partner: lay out the student’s goals and constraints, weigh the pathways (coursework vs research, local vs abroad, scholarship routes) in a simple matrix, and turn the best fit into a concrete timeline. You bring the wisdom; the structure makes the conversation productive.

Technique this week: decision-matrix prompting — you give the options and the criteria that matter to this student, and the model scores the trade-offs and drafts a roadmap. A tool for the conversation, not a verdict.

💡 Handy follow-ups: "Draft a scholarship-application timeline," "List programmes that fit these goals," "What should this student do in the next 30 days?" or "Give a Bahasa Melayu summary to share." Remember: a matrix organises a choice; it doesn’t make it — the human conversation is the point. Completes the student arc alongside Isu #18 (poster) and #23 (reference letters).

A mentor who actively encourages a student’s goals is rare — and it changes the trajectory of their whole career.

Gallup–Purdue Index (2014, 30,000+ US graduates): those who strongly agreed they had “a mentor who encouraged their goals and dreams” were ~2× as likely to be engaged at work and thriving; only ~22% strongly agreed they had one, and professors are ~two-thirds of college mentors. International benchmark.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The Methods section: the part everyone underestimates, then rewrites four times because Reviewer 2 “could not reproduce the procedure.” 😅 And if English isn’t your first language, it takes noticeably longer to get there. This week we use Gemini to expand your bullet-point protocol into a full, ordered Methods draft — measured against a reproducibility checklist so nothing essential is missing. You supply the facts and the accuracy; the AI handles the scaffolding and the English.

Technique this week: expansion-from-notes with a checklist constraint — you give terse bullets and an explicit “must-include” list, and the model writes to it, so the draft is complete, not just fluent.

💡 Handy follow-ups: "List everything I marked [TO CONFIRM]," "Tighten this to the journal’s word limit," "Add a data-availability statement," or "Check it against the journal’s reporting guideline." Fairness note: using AI for language support levels the field for non-native writers — that’s good scholarship, not cheating (disclose per policy). Feeds Isu #24 (reproducibility & data-management).

Writing in English costs non-native speakers real time and higher rejection — practical AI language support narrows a genuine gap.

Amano et al. (2023, PLOS Biology 21(7):e3002184): non-native English speakers need 51% more time to write and 91% more to read a paper, are rejected 2.6× more often, and asked to revise 12.5× more often for language. International benchmark; label as such.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Scholarship season arrives, and with it five requests for reference letters, each due “by tomorrow, Prof, so sorry.” 😅 Write them in a rush and they blur into the same vague praise — which quietly disadvantages the quieter students. This week we use Gemini to turn a student’s real achievements into a specific, evidence-based letter — then check it for the subtle language patterns research shows can undercut some candidates. Faster to write, fairer in effect, and still unmistakably your voice.

Technique this week: evidence-to-narrative prompting — you feed concrete achievements and the model shapes them into a specific letter, then audits its own language for fairness. Two moves in Gemini.

💡 Handy follow-ups: "Rewrite grindstone phrases as evidence of ability," "Make a shorter version for an online form," "Match the tone to a research scholarship," or "Draft a Bahasa Melayu version." Ethics first: a reference must be your honest assessment — AI structures the evidence, but the judgment and the signature are yours. Part of the student arc with Isu #18 and #21.

The same student can be described in ways that help or quietly hinder — evidence-based letters reduce an unconscious gap.

Dutt et al. (2016, Nature Geoscience): women ~half as likely to get an “excellent” vs “good” letter. Trix & Psenka (2003, 300+ letters): 34% of women’s letters used “grindstone” (hardworking-not-brilliant) terms vs 23% of men’s. International benchmark.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Reproducibility is like flossing: everyone agrees it matters, and somehow it never happens until there’s a problem. 😅 Then a grant demands a data-management plan (DMP) and you’re staring at a blank template at midnight. This week we scaffold it properly: point NotebookLM at the funder’s guidelines so it answers from the real rules, and use Gemini to turn your project into a DMP and a short reproducibility checklist. Build the habit once, and future-you is grateful.

Technique this week: template-scaffolding from an authoritative source — you ground the model in the funder’s own guidelines, so the plan follows their rules, not a generic guess.

💡 Handy follow-ups: "Turn the checklist into a README template," "What sensitive-data steps apply here?" "Summarise the funder’s sharing requirement," or "Draft a consent line for data reuse." Do it early: a DMP written at the proposal stage is painless; one written after data loss is impossible. Rigour thread: follows the Methods work in Isu #22.

Most researchers have hit reproducibility failures — a plan built up front is cheap insurance against a costly one.

Baker (2016, Nature 533:452; survey of 1,576 researchers): >70% have failed to reproduce another’s experiments, >50% their own, and ~90% agree there is a reproducibility crisis. International benchmark.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 The paper is done, the argument is sharp, and the reference list is quietly sabotaging you: a 2019 in the text that’s 2018 in the list, a citation with no entry, an entry no one cites. 😅 Reviewers notice, and it reads as carelessness. This week we run a consistency audit: have Gemini in Docs cross-check every in-text citation against the reference list, flag mismatches, missing entries and orphans, and check the formatting is uniform. Tedious for you; effortless for the machine.

Technique this week: consistency-audit prompting — you ask the model to cross-check one set against another and report only the discrepancies, so you fix instead of re-read.

💡 Handy follow-ups: "List only entries never cited in the text," "Check every DOI resolves," "Flag inconsistent author initials," or "Standardise all entries to APA 7th." Reality: AI can miss subtle slips — this narrows the haystack, but the final read is still yours. Closes the rigour thread begun in Isu #17 and #20.

Reference errors are far more common than we’d like to admit — and they’re exactly the kind of slip an audit catches.

Logan et al. (2025, J. Librarianship & Information Science): overall reference error rate ~21.2%. Onwuegbuzie et al. (2012): reference-list error rates ranged 4.1%–66.7% across studies. International benchmark.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Somewhere around interview transcript number twelve, every qualitative researcher has the same thought: “did I code that the same way in interview three?” 😅 Hand-coding is slow, and consistency quietly erodes as fatigue sets in. This week we use NotebookLM (grounded in your transcripts) to apply your codebook as a first pass and surface candidate themes — while you keep the part that makes it research: a real second coder and an agreement check. AI drafts the coding; you and a colleague own the interpretation.

Technique this week: deductive codebook-application — you define the codes up front and have the model apply them consistently to every transcript, then you run a human inter-coder check. Best in NotebookLM, grounded to your data.

💡 Handy follow-ups: "List every excerpt coded X with its quote," "Show passages that fit no code," "Summarise themes by participant group," or "Flag where two codes overlap." Rigour, not shortcuts: AI first-pass coding still needs a human second coder and a reported reliability figure — a shared codebook is what makes coding reproducible. Extends the rigour thread from Isu #20 and #24 into qualitative work.

Hand-coding is slow and fatiguing — and a shared codebook is what turns inconsistent coding into reliable coding.

Indicative figures from qualitative-methods sources (coding commonly ~4–6 hours per transcript; agreement declines with fatigue). A shared codebook matters: in one study (EUCA, 2021) inter-coder agreement rose from Cohen’s κ 0.43 to 0.88 after the codebook was refined. AI drafts a first pass; humans still verify and compute reliability.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 A student sends 60 pages and asks “is this okay?” — and it’s tempting to fix commas for an hour and call it supervision. 😅 But the feedback that actually helps is about structure: does the argument hold, do the chapters connect, is the research question answered? This week we use Gemini (or NotebookLM grounded in the draft) to run a structural diagnostic — a reverse-outline, gaps in the argument, weak transitions — so your comments target what matters. You still bring the disciplinary judgment; the tool clears the fog first.

Technique this week: structural-diagnostic prompting — you ask for the argument’s architecture, not a line edit, so feedback lands on what actually blocks progress.

💡 Handy follow-ups: "Rank these issues by how much they block completion," "Draft encouraging-but-honest wording for each," "Check the research question is answered in the conclusion," or "Give a Bahasa Melayu summary for the student." The point: structural feedback early beats endless late line-edits — and how you say it matters as much as what you say. Supervision arc with Isu #21 (mentoring).

Feedback is one of the highest-leverage things a supervisor gives — but only when it targets structure and substance, not surface.

Hattie, Visible Learning (2009): feedback d ≈ 0.73, among the top influences, vs the 0.40 “hinge point” (average influence). Wisniewski, Zierer & Hattie (2020, 435 studies, N>61,000): a more conservative d ≈ 0.48 — still above average. Note: feedback quality is decisive — vague or purely negative feedback can even backfire.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 You spend a year on a study, then thirty seconds picking a journal because it was the first one that came to mind. 😅 Weeks later: “we regret your manuscript is outside our scope.” A desk rejection before review is the most avoidable delay in publishing. This week we use Gemini — with NotebookLM grounded in candidate journals’ aims & scope — to match your paper to venues that actually fit, score the fit, and flag scope or predatory-journal risks before you submit. Aim before you shoot.

Technique this week: scope-fit matching — you describe your paper and its contribution, and the model scores candidate journals against their real aims & scope, so you submit where you fit.

💡 Handy follow-ups: "Draft a cover letter for the top pick," "Which of these are indexed in Scopus / WoS?" "Compare their scope in a table," or "What would make my paper fit journal #2?" Reality: AI can be wrong about a journal’s current scope or status — always confirm on the journal’s own site and a trusted whitelist. Research-writing arc with Isu #22 (methods) and #25 (references).

Pick the wrong journal and your paper can be rejected before it’s even read — scope mismatch is a leading cause.

Desk-rejection rates run ~30–70% at major journals (higher at the most selective). In one journal’s 2023 breakdown, ~35% of desk rejections were for being out of scope (vs ~7% for poor writing), and ~67% of early-career researchers get no feedback on a desk rejection. Indicative, field-dependent figures.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 December arrives, and with it the annual ritual of trying to remember what on earth you accomplished this year — buried across emails, calendars, and half-finished folders. 😅 The work was real; the record is chaos. This week we use Gemini in Workspace to harvest your year — from Calendar, Gmail and Drive — and shape it into an LNPT self-appraisal narrative mapped to the competency areas. You did the work; let the AI help you give it an honest, well-organised account. Just don’t let it inflate a single line.

Technique this week: achievement-synthesis — you gather a year of real activity and have the model organise it into a narrative against the appraisal’s own criteria. Uses Gemini across your Workspace.

💡 Handy follow-ups: "List all my 2026 emails about supervision," "Turn this into the LNPT form’s sections," "Draft a Bahasa Malaysia version in formal register," or "Suggest evidence I should attach." Honest and yours: a self-appraisal is a factual account, not a sales pitch — AI organises what you actually did; you confirm every claim. Year-end admin, sorted before the holidays.

Paperwork already eats a huge share of academic time — a tool that drafts your self-appraisal claws some of it back.

FDP Faculty Workload Survey (2012, 13,000+ principal investigators): researchers spend ~42% of their time on administration rather than research (consistent with the 2005 survey; a 2025 NASEM review puts it ~44% and no better). International benchmark; label as such.

Short version: the account you sign in with decides your protection — not the tool.

🌍 For the sovereignty-minded: "not trained on" isn't the same as "stored in Malaysia." True data-residency is a region setting your IT admin controls — worth a quick check if that's a hard requirement.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 It's the last Friday of the year, and if you're reading this instead of being on leave — thank you, and also, please go outside. 😄 Thirty Fridays ago, Kopi & Prompt set out to solve one faculty headache a week, one cup of kopi at a time. Since then we've drafted kertas kerja, screened abstracts, caught hallucinated citations, rehearsed vivas, mentored students, and untangled a Cronbach's alpha or two. Terima kasih for reading, forwarding, and actually trying the prompts. This week there's no new headache — just a look back, and a small gift waiting for you in the new year.

Salam sejahtera dan selamat pagi, rakan-rakan sekalian! 👋 Dan Selamat Tahun Baharu! We promised a gift, and here it is. Every Kopi & Prompt issue — all of them — now lives on a single webpage you can bookmark, search, and share with a new colleague. 🎉 And because this newsletter was always about doing it yourself, we're ending on the most on-brand prompt we could think of: one that builds a webpage. Paste it into Gemini and get a clean, single-file site for your own course, lab, or resources. One last cup — then go build something.

Technique to end on: one-shot web-page generation — describe the page you want and get a complete, single-file site back. Paste this into Gemini.

💡 Ask for edits: "add a search box," "make it dark mode," "add a table of my weekly topics," or "put a photo header." Then verify: open the file, click every link, and check it on your phone before you publish. The same rule as every issue this year — the AI drafts; you check and own the result.