On this page
TL;DR:
- The way teachers in Nairobi use AI well starts with real documents: the official timetable, past papers, marking schemes, KNEC feedback reports and your own learners' scripts. AI reorganises and drafts from those. It knows nothing reliable about this year's paper.
- The five jobs where it helps most are planning a revision block, drafting practice questions in the style of real papers, turning a tally of errors into a reteach plan, drafting parent updates and reviewing your own lessons.
- Solve every question yourself. Never call AI questions "KNEC papers" or "predicted papers".
- Learners' names and marks stay out of personal AI accounts, and one child's results never go in a parents' group.
- Knowing the learners, judging the marks and telling a parent the truth all stay with the teacher.
Who this guide is for
This guide is for tutors and subject teachers in Nairobi who prepare learners for national exams and have little time to do it. You may teach alone from a phone, or as part of a small team.
If you are a head or a department lead deciding AI rules for a whole school, start with our guide to AI for teachers in Kenya. It covers policy, learners' data, AI detectors and CBE lesson planning. This piece stays on the preparation work of exam-focused teaching.
A note on honesty. We have no survey of how teachers in Nairobi use AI, and as of 8 November 2026 we could not find one. So this is about what works. It draws on published research, on our own working methods, and on examples we have written ourselves and labelled as examples.
What makes teaching in Nairobi's exam season different?
Exam-driven teaching punishes small errors of fact, and that shapes how AI should be used.
| The reality | What it means for using AI |
|---|---|
| A revision group may bring together learners from several schools who have covered different parts of the syllabus | Ask learners what they have covered before AI plans anything |
| Time is short and blocks are fixed | A plan must fit the real hours, including the first ten minutes lost to settling in |
| KCSE papers are marked against a scheme | AI's general knowledge is the wrong standard. The scheme is the standard |
| Parents are paying attention, and often ask how their child is doing | Updates need to be specific and honest, and private |
| Many teachers work from a phone, on data that drops | Short prompts, short outputs and one job at a time |
That last row has some evidence behind it. In a one-day pilot with twelve Kenyan secondary teachers, run by an edtech company in May 2025, teachers said computers in public schools were few and shared, that phones were often the main way to reach tools, and that connectivity was unreliable. Twelve people is a small group, and the company has a product to sell.
Nairobi does give you one real advantage. The 2025 KCSE feedback reports, question papers and marking schemes were reported in September 2026 as on sale at the KNEC Bookshop in Nairobi. The feedback reports exist to show teachers where candidates were strong and weak.
Why real documents have to come first
AI gives you its best work when you hand it the source and tell it to stay there. Asked from memory about KCSE, it will produce something that sounds right.
We could not find any published test of how accurately AI tools answer or write KCSE questions, so assume nothing. Build a small source pack for each subject.
| Document | Where it comes from | What you use it for |
|---|---|---|
| Official timetable | KNEC | Every date and paper in your plan |
| Past papers | KNEC Bookshop or a trusted copy | Question format, command words, mark allocation |
| Marking schemes | KNEC Bookshop | The standard for model answers |
| KNEC feedback report | KNEC Bookshop | Where candidates actually lost marks |
| Your learners' scripts | Your own marking | What this group gets wrong |
| Syllabus or scheme of work | Your school or KICD | What has been covered and what is examinable |
A few habits follow. KNEC's name and copyright are printed on its papers, so use them for your own preparation and do not republish them. And take care with anything claiming to be a real paper: KNEC warned in September 2025 that papers circulating on social media were fake, and said official information is available only on its website.
The four Cs, in one minute
At FuKazee we teach four things to put in a prompt:
- Clarity: exactly what you want.
- Context: the subject, form, group and source material.
- Constraints: length, format and what to leave out.
- Customisation: who will read it and where it will be used.
Replace the square brackets in the prompts below. Never add a learner's name.
Planning a revision block
AI is useful for fitting topics into hours you have already counted. The counting, and the choice of topics, are yours.
Here is a mistake we made ourselves. During the 2026 KCSE, an AI assistant helping with our daily exam countdown labelled 2 November as Day 1 of the exams. It had counted only the main written block. Orals and practicals had begun on 19 October. Counting every weekday from then, and skipping the public holiday on 20 October, 2 November was Day 10. A person who knew the real timetable caught it.
News reports did not fully agree on the written start date either. The Star said in March that theory papers would start on 19 October, while the Ministry calendar as reported gave 2 November. Read the whole official timetable, including orals, practicals and projects, and work from that. If a count is a judgement call, write down how you counted.
The research on how to fill those hours is clear enough to plan around.
- Mix testing into every session. A major review of learning techniques rated practice testing and spreading study over time as the two most useful methods, and rated rereading, highlighting and summarising low.
- Mix topics in practice. In a randomised trial with US seventh-grade maths classes, students whose worksheets mixed problem types scored 61% on a surprise test a month later, against 38% for those practising one type at a time. That was American maths classes. Try it on your own group and see.
Prompt 1: fit a block to the real hours
I am preparing a revision block for [subject], [form or grade], with [number] learners from different schools. I have [number] sessions of [length]. Topics, in priority order, based on my own marking and the KNEC feedback report: [list]. Learners have not all covered: [list]. Draft a session-by-session plan. Each session must open with a 10-minute closed-book quiz on earlier topics, and include timed practice in exam format and mixed-topic questions. Bring every major topic back at least twice. Do not add topics or dates I have not given you. Mark anything you are unsure of [CHECK], because I will decide the final plan. Show it as a table a teacher can read on a phone.
Worked example (invented for this post). A tutor has five 90-minute sessions and a group of mixed Form 3s.
| Session | First 10 minutes | Main work | Last 20 minutes |
|---|---|---|---|
| 1 | Short diagnostic, no marks recorded | Priority topic A, worked examples | Five questions on topic A |
| 2 | Quiz on topic A | Priority topic B | Mixed questions, A and B |
| 3 | Quiz on A and B | Common errors from sessions 1 and 2 | Timed exam-format section |
| 4 | Quiz on B | Priority topic C | Mixed questions, A, B and C |
| 5 | Mixed quiz, all topics | Timed paper-style practice | Learners mark against a scheme and list two fixes |
For the longer picture, our realistic KCSE revision timetable is written for learners and useful to share.
Building practice questions in the style of real papers
AI can write many questions quickly if you show it what a real paper looks like. You then have to solve each one.
Copy the format exactly, because format is part of what you are practising. A meta-analysis of practice testing found practice tests generally beat restudying.
Take KCSE Mathematics Paper 1 from 2024 as an example. The cover says 2½ hours, two sections, all questions in Section I and any five in Section II. Candidates must show all steps, and marks may be given for correct working even if the answer is wrong. A worksheet that ignores those rules trains the wrong habits.
Prompt 2: questions in the style of a paper
Below are three real questions from a past [subject] paper, with their marks. Write [number] new questions in the same style and at the same level on [topic], for [form]. Keep the same command words, mark allocation and layout. Change the numbers, the context and the scenario. Use Kenyan settings where a setting is needed. Do not copy any question. For each one, give a full marking scheme showing where each mark is earned. Mark anything you are unsure of [CHECK]. I will solve every question myself before using it. Questions: [paste]
Prompt 3: test your own questions
Here is a practice question and my marking scheme. Act as a strict marker. List any ambiguity in the wording, any answer that could reasonably earn marks but is missing from the scheme, and any error in the working. Then write a typical weak answer and show how many marks it would earn and why. I will decide what to change. Question: [paste] Marking scheme: [paste]
Then do these checks, in order.
- Solve it yourself on paper, without looking at the AI's answer.
- Check it against the syllabus. Is this topic and level examinable for this form?
- Check the mark allocation against a real question of the same type.
- Draw any diagram, graph or apparatus yourself. Treat these as the items most likely to go wrong.
- Read for local sense. Prices, names and places should fit a Kenyan learner.
How good is AI at writing test questions? The best comparison we found is from medicine. In a 2023 study, ChatGPT wrote 50 questions in about 20 minutes, while two university staff took about 211 minutes. Independent assessors found no significant difference in overall quality, but the AI's questions scored lower on relevance and varied more. A KCSE paper is a different test, so treat that as a reason to save time and a reason to keep checking.
Label these as practice questions in your own name. Calling them "KNEC", "leaked", "predicted" or "likely to come" is untrue and can mislead families.
Spotting common errors
The best use of AI here starts with a tally you make by hand. It helps you decide what to reteach. It should not be marking the scripts.
Reading scripts is where you learn what the group does not understand, so keep that part. When our team filmed experienced KCSE examiners explaining why candidates lose marks, the same small things came up again and again in Biology graph work: the scale, the axis labels, where a line starts, and when to use a ruler. Those details differ by paper, which is why the rule must come from the marking scheme and from your own reading of the scripts.
Worked example (invented for this post). After marking 30 scripts on one graph question, a teacher writes this tally.
| Error | Scripts |
|---|---|
| Axis not labelled with units | 14 |
| Scale does not use most of the grid | 11 |
| Points plotted incorrectly | 6 |
| Curve joined by the wrong method for that question | 9 |
| Answered "describe" as if it said "explain" | 12 |
The tally holds no names, so it is safe to paste.
Prompt 4: from tally to reteach
After marking [number] [subject] scripts on [question type], this is my tally of errors: [paste tally]. Here is the marking scheme: [paste]. Rank the errors by likely marks lost. For the top two, suggest a 15-minute reteach that makes learners do something, and write three short practice tasks aimed at that exact error. Do not suggest going over the whole topic again. Use only chalkboard and exercise books. These are suggestions: I will choose what to teach.
Compare your tally with the KNEC feedback report for that paper. Where your group matches the national pattern, you have a strong case for spending the time. Where it differs, you have learned something about your group.
One boundary is firm. KNEC, as reported in September 2026, said teachers must not use AI to complete, or help learners complete, School-Based Assessment tasks. Everything in this section is about preparation.
If you do paste learners' work, remove the name and index number, and anything else that could identify them. Our AI for teachers guide has a traffic-light table for this.
Parent updates that are specific and honest
AI can give you a clear, warm draft in seconds. The facts in it have to come from you, and each message has to go to the right parent only.
Use a template with blanks. Fill in the child's name and details yourself, outside the AI tool. The ODPC's guidance for the education sector lists disclosure of academic records and other details to larger groups through parents' WhatsApp groups among its privacy concerns. Section 33 of the Data Protection Act also sets out rules on processing a child's data, including a parent's or guardian's consent.
So the group gets general notices. Results and concerns about a child go privately to that child's parent.
Prompt 5: a parent update template
Write a WhatsApp message template from a tutor to one parent, after [number] weeks of [subject] revision. Use these blanks: [CHILD], [ONE THING IMPROVED], [ONE THING STILL WEAK], [WHAT WE ARE DOING ABOUT IT], [ONE THING TO DO AT HOME]. Under 90 words. Warm, respectful and plain. Do not predict a grade or promise a result, and leave every blank for me to fill in. Give an English version and a Kiswahili version that reads naturally, not word for word.
Example of a finished message (invented for this post).
Good evening, Mama [name]. A quick update on [child]. She is now setting out her working step by step, which is earning her more marks. She still loses marks by not reading what the question asks. We are practising that at the start of each session. At home, please ask her to explain one question to you each evening. Thank you for your support.
A fluent colleague should read the Kiswahili before you use it. Parents can tell when a message could have been sent about anyone. Include one thing only you could know.
Never let AI soften bad news into something untrue. If a learner is not on track, the parent needs to hear that, with a plan.
Lesson-feedback routines
AI can act as a second pair of eyes on your teaching, and this may be its most underused job. Evidence on this is early, and it is promising.
- In a randomised trial with 900 tutors and 1,800 school students, students whose tutors had access to an AI suggestion tool were 4 percentage points more likely to master topics. The gain was 9 points for students of lower-rated tutors. Tutors also reported that some suggestions were not right for the grade level.
- In a randomised trial with 414 mentors, automated feedback on their talk increased how much they built on what students said by 10% and cut their talk time by 5%.
- A meta-analysis of debriefs across 46 samples, in medical and other settings rather than schools specifically, found that structured reflection improved performance by about 25% over control groups.
Those tools were purpose-built, and as far as the papers report, none of these trials took place in Kenya. What carries over is the habit: look at what actually happened in the lesson, then change one thing.
A 10-minute routine after a lesson
- Straight after the lesson, record a three-minute voice note on your phone, in your own words. What did you ask? Who answered? Where did learners get stuck? What did you do about it?
- Remove any learner's name from the transcript.
- Use the prompt below.
- Choose one change for the next lesson. Write it at the top of your plan.
Prompt 6: reflect on a lesson
Below are my own notes on a lesson I taught on [topic] for [form]. Do not praise me or give general advice. Tell me: how many of my questions needed reasoning and how many needed recall; where I answered my own question; where I might have built on what a learner said and did not; and which part of the lesson learners did the least thinking in. Then suggest one small change for the next lesson, in one sentence. If my notes do not give enough evidence for a point, say so. Notes: [paste]
If you want to record a whole lesson, get proper consent first, keep learners' voices out of any AI tool, and start with your own voice only.
Check any transcript you rely on. When we turned examiners' classroom talks into short videos, machine transcription of room audio could hear words that were never said, especially under music. We rechecked each finished video against what was actually said, because small errors change meaning. A lesson transcript deserves the same care.
What must stay with the teacher
Some parts of this job should never be handed over, however good the tools become.
| Stays with you | Why |
|---|---|
| Deciding what this group needs next | You have seen their faces and their scripts |
| Solving every question and checking every answer | One wrong scheme teaches a wrong method to everyone |
| Awarding marks | The judgement is professional, and it has to be yours |
| Every date, paper and count | These come only from the official timetable |
| What you tell a parent | Your name is on it and the child's future is in it |
| Noticing a learner who is struggling at home | No tool knows why a child has gone quiet |
| Saying "I do not know, let me check" | Learners need to see an adult do this |
| Keeping learners' work honest | Revision help must never become doing their assessed work |
What goes wrong, and how to avoid it
| Problem | What it looks like | What to do |
|---|---|---|
| Invented exam facts | A paper structure or date that sounds right | Check against the KNEC paper or timetable |
| Wrong marking scheme | A confident answer with a mistake at step three | Solve it yourself first |
| Overseas drift | Dollars, snow or another country's syllabus | Name the syllabus and paste the source |
| False precision | "Topic X comes every year" stated as fact | Count it yourself across real past papers |
| Over-promising | Materials or messages that imply a grade | Promise effort and a plan, never a grade |
| Learner data in a chat | A class list pasted in to "sort the marks" | Use a spreadsheet for marks |
| Reliance creep | A lesson taught without being read first | Read and change every draft before using it |
| Too much material | Forty new questions and no time to mark them | Ten good questions, then time for feedback |
A simple way to begin this week
- Monday. Buy or find the latest feedback report for your main subject and read one section.
- Tuesday. Run Prompt 2 on one topic. Solve every question.
- Wednesday. Use five of those questions in class. Write a tally of errors.
- Thursday. Run Prompt 4 on the tally and teach the reteach.
- Friday. Do the 10-minute routine and note what changed. Time the whole week, including checking, so you know honestly whether it saved you anything.
Questions people ask
Do teachers in Nairobi actually use AI?
We could not find a survey that answers this for Nairobi as of 8 November 2026. The closest evidence we opened is the twelve-teacher Kenyan pilot above. Assume the picture varies widely from one staffroom to the next.
Can AI predict what will come in KCSE?
No. It has no knowledge of an unreleased paper. You can count real topic frequency across past papers yourself, and even that is a guide to practice, never a promise. Be wary of anyone selling "predicted" or "leaked" papers.
Is it allowed to use AI to write practice questions?
We could not find a KNEC rule on using AI for lesson preparation, either way. The September 2026 warning, as reported, targets School-Based Assessment work, and the report describes it as no blanket ban on AI. Check any newer circulars.
Can I give AI a learner's script to mark?
Do not rely on it for marks. If you want suggested feedback, remove all identifying details first and use an account you trust. The final judgement is yours.
How do I use AI with poor internet and only a phone?
Use short prompts for one job at a time, and copy useful outputs into your notes while you have a connection. Voice typing helps. Paste only the piece of a paper that you need, and prepare the night before where you can.
Should I tell parents I use AI?
Yes, if they ask, and it is sensible to mention it anyway. Say that you use it for drafts and practice material, that you check everything yourself, and that their child's personal details stay out of it.
Which AI tool is best for teachers?
We do not promote one tool. In our experience, as of November 2026, the free plans of the main assistants can handle short drafting like this. Our AI for teachers guide compares the free routes and their data settings.
Sources
KNEC and exam information
- Kenyans.co.ke, KNEC releases 2025 KCSE and KJSEA reports and marking schemes (7 September 2026), last checked 8 November 2026
- The Standard, KCSE 2026 begins with oral and practical tests (19 October 2026), last checked 8 November 2026
- The Star, breakdown of 2026 KPSEA, KJSEA and KCSE timetables (4 March 2026), last checked 8 November 2026
- The Eastleigh Voice, KCSE to start 2 November as Ministry unveils 2026 calendar (13 October 2025), last checked 8 November 2026
- KNEC, KCSE 2024 Mathematics Paper 1, copy hosted by teacher.co.ke, last checked 8 November 2026
- Kenyans.co.ke, KNEC warning on fake exam papers (5 September 2025), last checked 8 November 2026
- Education News, KNEC warning on AI and School-Based Assessment (25 September 2026), last checked 8 November 2026
- Office Holidays, Mashujaa Day 2026, last checked 8 November 2026
Law and regulators
- Data Protection Act, No. 24 of 2019 (ODPC copy), last checked 8 November 2026
- ODPC, Guidance Note for the Education Sector, last checked 8 November 2026
Learning research
- Association for Psychological Science, summary of Dunlosky and others on learning techniques (2013), last checked 8 November 2026
- Rohrer, Dedrick, Hartwig and Cheung, A randomized controlled trial of interleaved mathematics practice, last checked 8 November 2026
- Adesope, Trevisan and Sundararajan, Rethinking the Use of Tests: A Meta-Analysis of Practice Testing, Review of Educational Research, 2017, last checked 8 November 2026
AI and teaching research
- Wang and others, Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise (arXiv), last checked 8 November 2026
- Demszky and Liu, M-Powering Teachers (EdWorkingPapers, May 2023), last checked 8 November 2026
- Tannenbaum and Cerasoli, Do Team and Individual Debriefs Enhance Performance? A Meta-Analysis, Human Factors, 2013, last checked 8 November 2026
- PLOS ONE, ChatGPT versus human in generating medical graduate exam multiple choice questions, 2023 (abstract record, Kabale University library), last checked 8 November 2026
- Avallain, TeacherMatic in Kenya: insights from a one-day pilot with educators (August 2025), last checked 8 November 2026
Where FuKazee fits in
FuKazee trains people in tech, the future of work and the future of careers, and we have trained more than 3,000 people. Teachers and schools are part of that work, and our page on AI training for education in Kenya explains how it runs.
We assume nobody in the room knows AI yet. We use no jargon, show the free version of every feature and never push one tool. We teach the four Cs first, keep academic integrity and responsible use at the centre, and say plainly that nobody becomes an expert in a few days.
If your team of teachers would like to practise this on your own subjects, papers and marking schemes, tell us how many you are and which subjects you teach. Email lead@fukazee.com
Want guided support instead?
Not sure which programme fits? Tell us who you are and we will point you to the right path: corporate AI training, the ScholarShift Accelerator for postgraduates, or KCSE study plans and tuition.
Rodney has run FuKazee's training since 2022 and still teaches inside every programme line: corporate sessions, the postgraduate cohort and the KCSE examiner webinars. First Class Honours in Journalism and Mass Communication, University of Nairobi.
