AI for Everyone study guide · Lesson 2 of 14
Learn by Showing
Zero-Shot Requests · One-Shot Examples · Few-Shot Boundaries · Iterate with Evidence
This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.
What you'll be able to do
- Choose zero-, one-, or few-shot prompting and improve examples through deliberate iteration.
Zero-Shot Requests
A zero-shot request gives instructions without a worked example. It is a sensible starting point for familiar tasks, quick exploration, and cases where the desired format is easy to describe. Zero-shot does not mean context-free: you can still provide source text, constraints, and a rubric.
Use the first result as a probe. If the model understands the task and the cost of an error is low, keep the simple prompt. If formatting drifts or a category is subtle, add an example instead of piling on vague adjectives. The smallest prompt that reliably works is easier to maintain.
Three ideas to retain:
- Zero-Shot Means No Demonstration: A zero-shot prompt asks the model to perform a task without showing a completed input-output example. Instructions and context can still be detailed.
- Simple Tasks Often Start Directly: For a familiar transformation such as turning notes into bullets, a clear instruction may be enough to discover whether examples are needed.
- The First Output Is a Probe: A low-risk first attempt reveals ambiguities in the task, missing context, and format drift. Its failures are information for the next prompt.
Key term: Zero-shot prompt means “instructions without a worked example.”
One-Shot Examples
One-shot prompting adds a single demonstration. It is especially useful when the desired output has an unusual shape, label scheme, or tone. Choose an example that reveals the rule rather than an easy case that would be obvious without it.
Separate the example from the real input with headings or delimiters. Keep its content different enough that the model must transfer the pattern rather than copy the answer. After the output, inspect whether the demonstrated feature—not accidental wording—was what carried over.
Three ideas to retain:
- One Example Can Define a Pattern: A worked input-output pair can demonstrate the expected labels, level of detail, or transformation more precisely than style adjectives.
- Examples Need Boundaries: Clear markers around EXAMPLE INPUT, EXAMPLE OUTPUT, and YOUR INPUT reduce confusion between the demonstration and the live task.
- Choose a Representative Case: An example should exercise the rule you care about. An unusually easy or exceptional case can teach the wrong boundary.
Key term: One-shot prompt means “instructions plus one demonstration.”
Few-Shot Boundaries
Few-shot prompting supplies several demonstrations. Its greatest value is not repetition; it is boundary coverage. Include ordinary cases, a difficult edge case, and—when useful—a near miss. The set can communicate which differences matter and which should be ignored.
Examples consume attention and context, so curate rather than accumulate. Research shows that demonstrations can help through format, label space, and input distribution even when their labels are not perfectly informative; that does not make correctness optional. Accurate, relevant, diverse examples remain the practical default.
Three ideas to retain:
- Variation Teaches Boundaries: Several examples can show a category's center, edge, and near miss. Diversity is more informative than repeating nearly identical easy cases.
- Examples Communicate More Than Labels: Demonstrations can reveal output format, available labels, and the distribution of inputs, not only a direct reasoning rule.
- Curate the Set: More examples use more context and can introduce contradictions. Keep examples relevant, accurate, and distinct enough to add information.
Key term: Few-shot prompt means “instructions plus several demonstrations.”
Iterate with Evidence
Prompting is closer to testing than spell-casting. Save the original task, collect a few representative inputs, and compare versions against the same rubric. Changing one major variable at a time makes it easier to learn why a result improved or regressed.
When an answer fails, diagnose the failure class: misunderstood task, missing context, wrong format, weak example, unsupported claim, or unsuitable tool. Rewrite the smallest responsible piece. A reusable prompt is not a frozen incantation; it is a versioned instruction that earns trust on known cases.
Three ideas to retain:
- Prompting Is Test-Driven: A prompt version is useful when it performs well on representative cases and a stated rubric, not merely when one output feels impressive.
- Change One Major Variable: Holding the test cases steady while changing one instruction or example makes the effect easier to interpret.
- Diagnose Before Rewriting: Different failures need different fixes. Missing evidence is not repaired by a more enthusiastic tone, and a wrong format may not need a new model.
Key term: Prompt regression means “a previously good case fails after a change.”
Sources
Key terms
- Zero-shot prompting
- Requesting a task without supplying a completed demonstration.
- Demonstration
- A worked example included in a prompt to show how an input should map to an output.
- Few-shot prompting
- Prompting with a small set of worked demonstrations.
- Prompt iteration
- A cycle of testing, diagnosing, revising, and comparing instructions.
Key facts
Simple Tasks Often Start Directly
For a familiar transformation such as turning notes into bullets, a clear instruction may be enough to discover whether examples are needed.
The First Output Is a Probe
A low-risk first attempt reveals ambiguities in the task, missing context, and format drift. Its failures are information for the next prompt.
Examples Need Boundaries
Clear markers around EXAMPLE INPUT, EXAMPLE OUTPUT, and YOUR INPUT reduce confusion between the demonstration and the live task.

Choose a Representative Case
An example should exercise the rule you care about. An unusually easy or exceptional case can teach the wrong boundary.

Examples Communicate More Than Labels
Demonstrations can reveal output format, available labels, and the distribution of inputs, not only a direct reasoning rule.
Curate the Set
More examples use more context and can introduce contradictions. Keep examples relevant, accurate, and distinct enough to add information.
Change One Major Variable
Holding the test cases steady while changing one instruction or example makes the effect easier to interpret.
Diagnose Before Rewriting
Different failures need different fixes. Missing evidence is not repaired by a more enthusiastic tone, and a wrong format may not need a new model.
Matched pairs
Each item on the left goes with the one on the right.
- Zero-shot prompt
- Instructions without a worked example
- Probe run
- A first attempt used to discover failure modes
- One-shot prompt
- Instructions plus one demonstration
- Representative example
- A case that exposes the intended rule
- Few-shot prompt
- Instructions plus several demonstrations
- Boundary example
- A case near the edge of a category
- Prompt regression
- A previously good case fails after a change
- Test set
- Representative inputs used across prompt versions
Step by step
Build a few-shot example set
- Name the distinction the examples must teach
- Add one ordinary positive case
- Add a useful edge case
- Add a near miss or contrasting case
- Check labels, format, and instructions for consistency
Prefer a small, coherent set over a long pile of redundant examples.
Run a prompt improvement cycle
- Save the current prompt and representative inputs
- Score outputs with a short rubric
- Name the dominant failure pattern
- Change one major instruction or example
- Retest and keep the version only if it improves the set
Record regressions as well as wins; a prompt can improve one case while harming another.
Self-test
Answer each question from memory, then check your work against the answer key.
1. What can a zero-shot prompt include?
- AOnly one sentence
- BContext and constraints
- CA trained adapter
- DNo source material
2. When is zero-shot a good start?
- AA hidden grading label
- BA safety-critical decision
- CA novel private code
- DA familiar low-risk task
3. What is one-shot prompting?
- ATraining one parameter
- BRunning the model once
- CGiving one worked example
- DUsing one-word answers
4. Why delimit an example?
- ATo separate it from live input
- BTo guarantee accuracy
- CTo reduce all token use
- DTo add hidden memory
5. What makes a few-shot set useful?
- AVaried boundary coverage
- BHidden instructions
- CMaximum repetition
- DRandom long passages
Go deeper
- WatchThe First Prompt Is a Probe Hand a model your notes with one instruction: turn these into bullets. No example of the answer.
- StudyAI for Everyone in the FlashFeed app
Answer key
- 1. B Context and constraintsThe term concerns demonstrations, not the amount of instruction.
- 2. D A familiar low-risk taskSimple low-risk tasks let you test the direct request cheaply.
- 3. C Giving one worked exampleThe “shot” is a demonstration, not an execution count.
- 4. A To separate it from live inputBoundaries clarify which text is instruction, example, and task.
- 5. A Varied boundary coverageDistinct cases teach more of the decision boundary.
Sources
- Min et al. — Rethinking the Role of Demonstrations — arxiv.org/abs/2202.12837
- Google Cloud — Prompt design strategies — cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/prompt-design-strategies
Image credits
- Choose a Representative Case: AI-generated (gpt-image-1.5)
- Examples Communicate More Than Labels: AI-generated (gpt-image-1.5)