WEBVTT

NOTE
Subtitles for the Prompt engineering walkthrough. They match the captions drawn on the video,
and there is no spoken narration. The music is an original score. Cue times follow
the film timeline in seconds.

1
00:00:00.400 --> 00:00:04.870
Why does one model give very different answers to slightly different wording?

2
00:00:05.070 --> 00:00:08.490
The promise: a repeatable method beats trial and error.

3
00:00:08.690 --> 00:00:12.870
Our job: build a prompt from parts, test it, and make it safe to reuse.

4
00:00:13.900 --> 00:00:19.840
First, what a model actually sees. A system message with rules, role and tone, set by whoever built the app.

5
00:00:20.040 --> 00:00:24.700
Then your prompt, and the input data: the document, email or question to work on.

6
00:00:24.900 --> 00:00:30.370
The model predicts the next token, one at a time, and it only sees the text in its context window.

7
00:00:30.570 --> 00:00:36.510
Wording shifts the odds of each token, and sampling adds randomness, so small changes can change the answer.

8
00:00:37.900 --> 00:00:42.800
A good prompt has six parts. Start with the task: one clear job, starting with a verb.

9
00:00:43.000 --> 00:00:47.410
Add the context the model cannot know, and say who answers or who it is for.

10
00:00:47.610 --> 00:00:52.370
Set measurable limits and the output format, and show one example of what you want.

11
00:00:52.570 --> 00:00:56.320
More words is not the goal. Clear, checkable instructions are.

12
00:00:57.400 --> 00:01:03.440
In the workbench, watch the techniques first: be specific and measurable, and separate instructions from data.

13
00:01:03.640 --> 00:01:07.910
Show one example, name the output format, and change one thing at a time.

14
00:01:08.110 --> 00:01:14.530
Then the risks: injection from untrusted text, confident answers that are wrong, and secrets pasted into hosted tools.

15
00:01:14.730 --> 00:01:18.870
Predict: how many of the six parts does your last real prompt contain?

16
00:01:19.900 --> 00:01:25.560
This is the prompt workbench. Paste a prompt you already use, and the checker looks for the six parts.

17
00:01:25.760 --> 00:01:29.220
Build a prompt from parts, and see the assembled result.

18
00:01:29.420 --> 00:01:32.460
Compare weak and strong prompts in the gallery.

19
00:01:32.660 --> 00:01:37.600
And try a prompt injection safely, to see why instructions and data must be kept apart.

20
00:01:38.900 --> 00:01:44.840
Now the techniques that work. Zero-shot is instructions only, for simple tasks, though the format may drift.

21
00:01:45.040 --> 00:01:50.510
Few-shot shows the pattern with examples, but keep them varied. Step by step gives room to reason.

22
00:01:50.710 --> 00:01:56.940
Structured output fixes the shape of the reply, so always validate it. Chaining splits a big job into small steps.

23
00:01:57.140 --> 00:02:01.600
Self-consistency samples several answers and votes, which costs several runs.

24
00:02:02.900 --> 00:02:07.080
Prompting is iteration. Write a version, then test it on varied inputs.

25
00:02:07.280 --> 00:02:12.220
Compare the results with what a good answer looks like, and change one thing at a time.

26
00:02:12.420 --> 00:02:17.130
Record the version, the change and the result, and re-test when you switch models.

27
00:02:18.400 --> 00:02:23.390
Prompts and safety. Prompt injection means text in the data overrides your instructions.

28
00:02:23.590 --> 00:02:29.720
Separate instructions from data, limit what the model can do, validate the output, and keep a human in the loop.

29
00:02:29.920 --> 00:02:35.440
Models also hallucinate, so ask for answers drawn from text you supply and verify important claims.

30
00:02:35.640 --> 00:02:41.580
Redact secrets or use a local model, and never put secrets in a prompt. A prompt is not a security boundary.

31
00:02:42.900 --> 00:02:46.410
Here is the checklist for writing and running any prompt.

32
00:02:46.610 --> 00:02:52.130
Is the task one clear job? Did I give context the model cannot know, and say who the answer is for?

33
00:02:52.330 --> 00:02:57.460
Did I set measurable limits and show an example? Did I separate instructions from the data?

34
00:02:57.660 --> 00:03:02.600
Did I test on varied and tricky inputs? And did I check the facts and keep secrets out?

35
00:03:03.900 --> 00:03:06.900
Two questions for the room.

36
00:03:07.100 --> 00:03:13.090
Pick a task you repeat. Turn your prompt into a template: which parts change each time, and which stay fixed?

37
00:03:13.290 --> 00:03:17.900
If a prompt says never reveal the password, what else must protect the password?

38
00:03:18.900 --> 00:03:23.600
That is the lesson: build a prompt from parts, test it, and make it safe to reuse.

39
00:03:23.800 --> 00:03:26.800
A repeatable method beats trial and error.

40
00:03:27.000 --> 00:03:30.000
Clear, checkable instructions are the goal.