The words spoken in the Unit 5 video, scene by scene, as its captions show them, with what the animations show marked [On screen: ...] where they play. Each speaker is named where the speaker changes. The Unit 5 reading covers the same material.
The narration is spoken by synthetic voices made for this course, each shown on screen as a drawn robot. No person is speaking.
Unit 5 — Prompts, context and attachments
Narrator: You ask something and get a good answer. A colleague asks the same thing and gets something different. Neither of you did anything wrong.
What goes in
Sampling or a changed service can do that, as Unit 3 showed. So can something easier to miss: the two questions may not have been the same question, because a prompt is not what reaches the system.
[On screen: Headed 'The window does not grow', a fixed frame labelled context window. Pieces drop into it one at a time and stack up: standing instructions, a thin band labelled turn 1: your prompt, turn 1: your attachment, then turn 2, under the captions: everything the system sees, in order; attachments go in, as far as they fit; each turn can carry earlier turns. The thin prompt band is outlined and marked 'one part of it'. Then: the window is fixed. Turn 3 arrives under the caption: the next piece does not fit, something has to go. The standing instructions stay at the top; turn 1's prompt slides out into a list headed no longer in the window, marked 'turn 1 prompt: gone', and turn 4 pushes turn 1's attachment out after it, marked 'turn 1 attachment: gone'. Closing line: the earliest turn stops being available, and you are not told.]
For a text-only model, the context is a single assembled sequence, put together just before the answer is generated. Your typed prompt is one part of it: the thin band, outlined here. And the whole frame goes in together. The answer is generated from all of it, not from your prompt alone.
[On screen: Headed 'The window does not grow', a fixed frame labelled context window. Pieces drop into it one at a time and stack up: standing instructions, a thin band labelled turn 1: your prompt, turn 1: your attachment, then turn 2, under the captions: everything the system sees, in order; attachments go in, as far as they fit; each turn can carry earlier turns. The thin prompt band is outlined and marked 'one part of it'. Then: the window is fixed. Turn 3 arrives under the caption: the next piece does not fit, something has to go. The standing instructions stay at the top; turn 1's prompt slides out into a list headed no longer in the window, marked 'turn 1 prompt: gone', and turn 4 pushes turn 1's attachment out after it, marked 'turn 1 attachment: gone'. Closing line: the earliest turn stops being available, and you are not told.]
So besides your prompt, the context can include standing instructions the product adds, which you do not normally see; earlier turns of the conversation; attachments, converted to text or otherwise encoded; passages retrieved by a search; tool results; and saved memory carried over from earlier conversations. A multimodal system may also receive encoded images, audio or video. Look at where your prompt sits in that list. So good prompting is not a matter of finding magic words. Much of what shapes an answer is the other material that goes in with your prompt.
The window has an edge
[On screen: Headed 'The window does not grow', a fixed frame labelled context window. Pieces drop into it one at a time and stack up: standing instructions, a thin band labelled turn 1: your prompt, turn 1: your attachment, then turn 2, under the captions: everything the system sees, in order; attachments go in, as far as they fit; each turn can carry earlier turns. The thin prompt band is outlined and marked 'one part of it'. Then: the window is fixed. Turn 3 arrives under the caption: the next piece does not fit, something has to go. The standing instructions stay at the top; turn 1's prompt slides out into a list headed no longer in the window, marked 'turn 1 prompt: gone', and turn 4 pushes turn 1's attachment out after it, marked 'turn 1 attachment: gone'. Closing line: the earliest turn stops being available, and you are not told.]
The context is finite. When a conversation or a document exceeds it, something is dropped, or summarised. And overflow is not the only way material goes missing. A long document may be cut into pieces of which only some are retrieved. And one that fits whole may still be used unevenly: in one study published in twenty twenty-four, testing models from twenty twenty-three, material in the middle of a long input was used worst. None of it is usually shown to you. All of it means material you supplied may not be there when the answer is generated, or may be there and carry little weight.
A confident answer from part of the document
The system answers as though your document said less than it does — and the answer reads as confident, because the context it was generated from did say less. Suppose you attach forty pages and ask about a clause on page thirty-one, and page thirty-one did not survive the window, or was there and barely used. What comes back can still be fluent and wrong, built from the pages that did arrive, and cite one of them: page four. Nothing in its tone separates that from the case where the whole document was present. That is why a locator matters, the part of a citation that says where to look: it lets you check whether the cited passage exists and supports the answer. It does not tell you whether that passage reached the model. So for anything that matters, ask about one short part of a document, not the whole of it.
Attachments are not references
Attaching a document moves the task into grounded answering. The reference is now something you hold, so the check has a clear limit. What it does not do is make the answer an extract. What comes back is still generated text about the document. So: ask for the passage. Read the passage in the document, not in the answer. Then ask whether that passage settles the question you asked. The third one is the one people skip. A passage can be genuinely present, genuinely quoted, and about something adjacent.
Attached text can give instructions
Text inside an attachment or a retrieved page joins the same context as your instructions, and it can act as instructions: a line addressed to the system, visible or hidden, which the output then follows. This is called prompt injection. It was shown working against a real deployed system in twenty twenty-three. Hidden text counts too, and security guidance from twenty twenty-five lists no fool-proof prevention. So treat attached content as data, and check what the output did with it. If an answer does something you did not ask for, look in the source for the text that asked for it.
Before
Now one before and after, on a task with no code. A twenty-page handbook is attached, and you want to know what it says about late submission. This is the prompt as first written, the before version. What does this say about late submission? Fluent, plausible, possibly right. The problem is not that it is wrong. It is that you cannot tell, and finding out means reading the handbook, which was the work you were trying to save.
After
[On screen: A prompt card holds the vague prompt 'What does this say about late submission?'. It moves up to a line labelled 'before', and the card, now labelled 'after', fills with the specified prompt one sentence at a time: 'Using only the attached handbook, list every rule it states about late submission.' 'For each one give the section number and quote the sentence it comes from.' 'If the handbook does not state a rule about something, say so rather than filling the gap.' A line reads 'Three changes, none of them a magic word.' Each sentence is then bracketed and labelled in turn: 1, what would count as correct, with 'Using only the attached handbook' and 'every rule' picked out; 2, locators, with 'the section number' and 'quote the sentence' picked out; 3, what to do about absence, with 'If the handbook does not state a rule' and 'say so' picked out.]
Using only the attached handbook, list every rule it states about late submission. For each one give the section number and quote the sentence it comes from. If the handbook does not state a rule about something, say so rather than filling the gap. Three changes, none of them a magic word. It says what would count as correct. It asks for locators. And it says what to do about absence. Silence and "no rule exists" are different findings. If a request does not tell them apart, the answer tends to be whichever reads more fluently.
What did not change
Questioner: Is that answer verified?
Narrator: The second answer is not verified. It is checkable, which is different, and is the whole point. Two minutes confirming three section numbers, instead of an afternoon finding out whether each returned rule is right. That shows support for the rules returned. But completeness is separate. It still needs a search of the handbook for missing rules, for exceptions, and for clauses that say which rule wins.
Three things that are not magic words
Three things that reliably change an answer, and none of them is a phrasing trick. Stating what would count as correct is the biggest one. It changes the answer because it changes what you can compare the answer against. Giving an example of the form you want is cheap and useful, with one hazard. An example constrains the shape of the answer, not its truth — and a response in exactly the format you demonstrated can be entirely wrong, while the matching format makes it read as compliant.
Checker: One of the course's lab examples returns every answer in exactly the form its instructions ask for, an exact fraction. Compared with the known correct values, none of its thirty-one answers is right.
Narrator: And starting a new conversation is underrated. If an exchange has gone somewhere you did not intend, the accumulated context is now working against you, and clearing it beats adding more turns to correct it. None of the three makes the output self-verifying.
Three habits
Record the whole input. A prompt without its conversation is not a record of what produced the output. And if you keep it as a shared link, one provider's documentation, as of November twenty twenty-six, says anyone who has the link can read it. Copies made from it stay readable after you withdraw the link. Ask again in a clean context, not in the same one — inside the same conversation, the earlier answer is now part of the input. A new conversation is not automatically clean either. The same provider describes saved memory, account-wide instructions and project instructions that carry into new conversations, with memory able to be on by default. Note whether they were on. Ask differently, too. Rephrase the question, or reorder its options. Sometimes the right answer should not depend on the wording. If the answer then changes, one of the answers was wrong. You found that out without knowing the right answer. Testers call this a metamorphic relation. An answer that stays the same is consistent. That does not make it correct. And ask for locators every time. They cost the system very little, and when one turns out not to exist, you have learned something the fluent version would have hidden.
Next
The activity asks you to find a late clause in the last quarter of a long document, ask what only it can answer, and see what comes back, with the locator, and then with the document open in front of you. There are four possible outcomes and every one of them tells you something.
Closing, 20 seconds, no speech. A candidate record on the left, a check against an independent reference in the middle, and a bracketed claim on the right fade in side by side. Below them: Continue with the course material, ai.qyber.dev/edu-ai-literacy-ada-verification/, Course playlist: youtube.com/playlist?list=PLPsBWTfgvakM, All links are also in the description, and Keep your record. Check the next claim.