Lesson 1961 of 2116
Few-Shot Example Curation: Quality, Rotation, and Counter-Examples, Part 2
Negative examples sharpen behavior more than positive ones alone.
Lesson map
What this lesson covers
Learning path
The main moves in order
- 1The premise
- 2Anchor Trap: Why Your First AI Example Skews Every Output
- 3The premise
- 4Counter-Example Method: Show AI What 'Wrong' Looks Like
Concept cluster
Terms to connect while reading
Section 1
The premise
A good example tells the model where to aim; a bad example tells it where the cliff is. Pair them.
What AI does well here
- Match the style of provided good examples.
- Avoid patterns explicitly marked as bad.
What AI cannot do
- Generalize from a single example reliably.
- Infer the rule when good and bad look identical.
Key terms in this lesson
Section 2
Anchor Trap: Why Your First AI Example Skews Every Output
Section 3
The premise
AI weights early examples heavily. If your first sample is verbose, every output gets verbose; if it's casual, all become casual.
What AI does well here
- Match the tone, format, and length of your first example.
- Repeat structural patterns from samples you provide.
- Generalize from 2-3 well-chosen examples.
- Detect implicit rules in your demonstrations.
What AI cannot do
- Ignore a misleading first example you didn't intend as a pattern.
- Average out an outlier sample you put in by accident.
Section 4
Counter-Example Method: Show AI What 'Wrong' Looks Like
Section 5
The premise
Showing only good examples teaches the surface; showing good-vs-bad pairs teaches the boundary.
What AI does well here
- Learn distinctions from contrasting pairs.
- Avoid features marked as bad when generating new outputs.
- Match the 'good' side reliably after seeing the contrast.
- Calibrate edge cases when you label them.
What AI cannot do
- Generalize correctly from a single good/bad pair.
- Distinguish stylistic preference from objective error without your labels.
Section 6
Instructions vs Examples: When AI Needs Each
Section 7
The premise
Instructions work for tasks AI already knows; examples are needed for novel formats or rare patterns.
What AI does well here
- Follow clear instructions for known tasks (translation, summarization).
- Generalize from 2-3 examples for unfamiliar formats.
- Combine both when neither alone suffices.
- Tell you which approach a task likely needs if you ask.
What AI cannot do
- Reliably learn novel formats from instructions alone.
- Generalize from a single example without context.
Section 8
AI Few-Shot Example Selection: Which Examples Actually Teach the Model
Section 9
The premise
Few-shot example quality matters more than quantity — diverse, edge-case examples shape AI behavior, while redundant examples waste context with diminishing returns.
What AI does well here
- Generalizing from 3-5 well-chosen examples
- Mimicking format, tone, and structure shown in examples
- Handling edge cases similar to those demonstrated
- Following negative examples when explicitly framed
What AI cannot do
- Generalize beyond the distribution shown in examples
- Detect that an example set is unrepresentative
Section 10
AI Zero-Shot Classification: When You Can Skip Few-Shot Examples
Section 11
The premise
AI zero-shot classification works well when label names carry rich semantics, label sets are small, and classes are clearly distinct — degrading on close-call boundaries.
What AI does well here
- Classifying with semantically rich labels like 'urgent customer complaint'
- Handling clearly-distinct categories without examples
- Producing probability-like confidence when prompted
- Following definitions provided per label
What AI cannot do
- Distinguish between near-synonym labels reliably
- Improve indefinitely without examples on hard boundaries
Tutor
Curious about “Few-Shot Example Curation: Quality, Rotation, and Counter-Examples, Part 2”?
Ask anything about this lesson. I’ll answer using just what you’re reading — short, friendly, grounded.
Progress saved locally in this browser. Sign in to sync across devices.
Related lessons
Keep going
Creators · 40 min
Few-Shot Example Curation: Quality, Rotation, and Counter-Examples, Part 1
Chain-of-thought prompts show real performance gains on reasoning tasks — and zero benefit on tasks that don't need reasoning. Here's how to tell which is which.
Creators · 40 min
Meta-Prompting and Self-Critique: AI That Improves Its Own Output
Static templates are predictable and cheap. Generated prompts adapt to context. The decision shapes maintenance burden, quality, and team workflow.
Creators · 40 min
Prompt Security: Injection Defense, Jailbreaks, and Refusal Design
Prompt injection isn't solvable by prompting alone. Layered defenses combine prompt design, input filtering, and output validation.