Five untimed CELPIP Speaking Task 4 questions, each with its own picture and a spoken model answer. Every scene is frozen a second before something lands, and your job is to say what happens next. This is the task candidates lose most often, and the reason is always the same one.
On the real test, Task 4 gives you back the illustration you have just spent 60 seconds describing. Same picture, no changes, and one new instruction: what do you think will most probably happen next. Thirty seconds to prepare, 60 seconds to record.
The five questions below come with pictures of their own. On the exam, one illustration serves Task 3 and Task 4 between them. In this course the two tasks have separate scenes, deliberately, so that you get ten frames of practice rather than five: here it is a self-serve car wash, a ferry terminal, an off-leash dog park, a bank branch and a neighbourhood playground. Each one is built the way a CELPIP scene is built, with several people caught mid action and at least one thing nobody in the frame has noticed yet.
Now the failure, and it deserves to be stated bluntly, because it accounts for more lost marks than every vocabulary gap on this task combined. Candidates describe the picture again. They open with "in this picture I can see a car wash with several bays and a queue of cars", spend 40 of their 60 seconds on the same material the rater heard a minute ago, and squeeze one prediction into the final ten. That answer scores close to nothing on this task, not because it is bad English, but because it answers the previous question. The rater already has your description. The marks are exclusively in the sentences that leave the frame.
Here is the test you can run on your own recording: how many verbs are in the future or the conditional? In a strong Task 4 answer, almost all of them. Is going to, will most likely, would probably, gets soaked the second she straightens up, I doubt that bag stays on the bumper. If you are hearing is and are, you are describing, and describing has already been marked.
Four predictions all stated with will is four predictions with no ranking. Grading them by how sure you are costs nothing and shows range, precision and judgement at once.
There is a failure mode at each end of that ladder. Live entirely at the top and every outcome is a flat certainty, which sounds confident and reads as a speaker who cannot distinguish a wobbling scooter from a toddler walking into the path of a swing. Live entirely at the bottom and you hedge so heavily, maybe, perhaps, it could be, that no prediction is ever actually made. That second one is worse: Task Fulfillment scores relevance and completeness, and an answer that refuses to commit has not done the task at all.
The other move worth stealing is the chain. Do not leave your predictions as a row of unrelated sentences. Link two: the boy pulls the wand, so the jet swings wide of the hatchback, so the woman polishing her headlight in the next bay straightens up into a face full of cold water. A rater hearing a causal chain hears somebody who read the frame rather than scanned it, and it costs you no extra time at all.
The 2026 CELPIP speaking format for the prediction task specifically.
Once you have stopped re-describing, five things separate a competent prediction from a high one.
It is worth understanding why these scenes feel so full of imminent mishaps. A picture that works for prediction has to contain unfinished business: people caught mid transaction, somebody who has not noticed something, a party who must respond. A frame of people sitting quietly on benches would be a fine Task 3 and an impossible Task 4, because there would be nothing to predict but more sitting. So the scenes are engineered with mild momentum in them, and that is your material.
Mild is the operative word, though. The frames are built so that a careful reader and a wild one arrive at different answers. If your prediction is that the boy hanging off the climbing frame falls and breaks his arm, you have stopped reading a picture where his mother is already reaching up for his waist and started writing a screenplay. Everyone can predict a catastrophe. Predicting that the man wrestling with the stroller finds his runaway orange fifteen minutes later, covered in sand, and simply leaves it there is what a rater actually has something to reward.
Five illustrations, each holding more unfinished business than you have 60 seconds to use. Open one, read the frame, record your predictions with no clock running, listen back, then reveal the spoken model answer and the reasoning underneath it. Your recording stays in your browser: it is never uploaded and nothing here is scored.
A pressure wand arcing over a roof with a child hanging off it, a floor mat being shaken at a freshly washed car, an ice cream cone dangling from a van window. One arc here lands on somebody whose back is turned, and following it there is what makes an answer a chain rather than a list.
Practise this question
A barrier half closed, a family still running for it, and a duffel bag balanced on a bumper that is about to move. The gate is a genuine coin toss, so give both futures and then commit to one with a reason attached.
Practise this question
A large dog up on its back legs at the gate, a sandwich being reached for without looking, a lead looped around a wrist that is not paying attention. Most of these outcomes are settled by physics rather than by anyone's choice, and physics is a stronger reason than a guess.
Practise this question
A courier reversing through a glass door, a student on his phone who will not feel it coming, a boy being steered away from the candy bowl. The people here all have reasons, and predicting the reason scores higher than predicting the collision.
Practise this question
A toddler wandering into the path of a swing, three impatient children behind a girl at the top of the slide, an orange rolling quietly toward the sandbox. Rank them by how soon they land and forecast the most imminent one first, then attach the motive: she goes down that slide, but not because she wants to.
Practise this question"The first thing that is going to happen is ..." Make the very first verb of your answer a prediction and you have structurally locked yourself out of re-describing. It is a one sentence fix for the biggest problem on this task.
More is about to happen in that frame than you can fit into 60 seconds. Find the arcs in your 30 seconds and rank them by how soon they land. That ranking is your running order, and it means you never arrive at second forty wondering what else there is.
An orange rolling toward a sandbox. A duffel bag on the lip of an open trunk. A vacuum hose that has half swallowed a floor mat. Every planted detail that one person in the frame cannot see is a prediction handed to you for free, and it is the easiest mark on the task.
"If that cone comes any further out of the window, it is going down their own paintwork." One genuine conditional gives you a complex structure under full control and a consequence that nothing in the frame actually shows. Two is better. Zero is a missed criterion.
On test day the two tasks share one picture, and it is not a budget to split. Describe fully in Task 3, then predict fully here. Holding back detail in the description to have something to say later costs you marks twice.
"My guess is he finds it fifteen minutes later, covered in sand, and just leaves it there." A joke is fine. A joke that is still forward looking and still has a timeframe on it is better, and it keeps the last five seconds working.
CELPIP Speaking Task 4 is Making Predictions. The illustration you have just described in Task 3 comes back on the screen, and you say what will most probably happen next. You get 30 seconds to prepare and 60 seconds to speak. It is the fourth of the eight speaking tasks in the 2026 CELPIP format.
On the real test, yes. One scene illustration serves both tasks: you describe it in Task 3, then the same image returns immediately in Task 4 and you predict forward from it. Knowing that in advance matters, because anything you spend on prediction during Task 3 is material you will not have here. This practice course is deliberately different. Task 4 has five pictures of its own, so the two tasks give you ten scenes to work on instead of five.
Because they describe the picture again. It is the single commonest failure on this task and it scores close to nothing, since the rater already has your description from the previous 60 seconds. The marks are in the sentences that leave the frame. If your answer opens with in the picture I can see, it has already gone wrong.
Three or four, from different parts of the frame. A CELPIP scene is built with at least four separate arcs in it, each one a small unfinished story. Predicting one thing in ten seconds and stopping leaves two thirds of the task unused, and covering only the funniest outcome ignores the rest of the picture.
No, and this is where the easy marks are. Might, is going to, will probably, would most likely, almost certainly and I doubt all carry different degrees of confidence. Grading your predictions by how sure you are is a top band marker under both Vocabulary and Content. Stating everything as a flat certainty flattens the whole answer.
Every prediction has to be traceable to something visible. A fire alarm going off or an ambulance arriving is invention, not prediction, and it reads as an answer that could have been given about any picture. The ice cream cone hanging out of the van window, the orange rolling toward the sandbox and the duffel bag balanced on the open bumper are all planted for you: use them.
No. A well built scene carries mild momentum rather than catastrophe: somebody getting soaked, a queue backing up, a bag of chips lost to a seagull. If you turn it into a disaster, with a child falling off the climbing frame, you have stopped reading the frame and started writing fiction. Likely, not dramatic, is the standard.
Yes, and it is one of the strongest moves available. Predicting a motive rather than only an action, such as she goes down the slide, but not because she wants to, is abstract language sitting on top of a concrete guess. Most candidates predict only objects, and their answers stay flat as a result.
No. There is no timer and no evaluation anywhere in the practice course. You record yourself so that you can hear yourself, the recording stays in your browser and is never uploaded, and then you reveal a spoken model answer with its structure, key phrases and common mistakes.
Yes. Everything here follows the 2026 CELPIP speaking format: eight tasks, a timer that runs per task and resets on every advance, and Task 4 running 30 seconds of preparation followed by 60 seconds of recorded speech, using the illustration carried over from Task 3.
Task 5 gives you a full minute of preparation and somebody to argue with. Or go back and get the description right first.
Three screens, a rejected option that vanishes, and a third one you never saw coming.
Learn Task 5 → 3On test day, the task this one inherits its picture from. Get the route and the verbs right and both tasks improve.
Learn Task 3 →Every task, every question, every model answer, on one page with the topics listed in full.
See the samples →All 8 tasks back to back with the real prep and response clocks running, and real illustrations.
See the speaking section →Five pictures, no clock, no score. Record your predictions, then count how many of your verbs pointed forward. That number is the task.
