Better Prompting LLMs Through Analogies

cvs2681 pts0 comments

Better Prompting LLMs Through Analogy

How to better prompt LLMs

Make the next token<br>easier to choose

A good prompt<br>reduces ambiguity<br>reduces conversion work<br>and makes the desired operation match the model's learned patterns.

Accurate<br>Clear inputs and success criteria<br>produce fewer wrong branches.

Fast<br>Less hidden reasoning<br>means the model reaches the answer sooner.

Fewer Tokens<br>Compact structure<br>avoids repeated clarification<br>and repair.

Lower Cost<br>Shorter inference<br>and fewer retries<br>cut usage spend.

The practical model

Every unnecessary transform<br>is a chance to be<br>slower, longer, or wrong

The following fruit game will make it obvious.<br>You will sort the same fruits three ways<br>and compare the cognitive overhead.

See the input<br>The model receives tokens<br>images<br>code<br>logs<br>instructions.

Map it<br>It converts the request<br>into a familiar learned pattern.

Choose action<br>It follows constraints<br>selects tools<br>or generates an answer.

Emit output<br>The requested answer shape<br>determines how much cleanup is needed.

Attempt 1

Image in<br>string labels out

The fruit moves toward you.<br>The boxes show text labels.<br>Press an arrow key<br>or click the matching box<br>before the fruit reaches the boxes.

Task: match a fruit image to a word<br>then choose the matching side.

Hidden cost: image-to-label conversion<br>happens on every fruit.

Run length: 20 fruits.

Fruit0 / 20

Correct0

Avg All0.00s

Avg Correct0.00s

left

right

Sort 20 fruits

Click a box<br>or use the keyboard arrows<br>to sort each fruit.

Start attempt 1

Attempt complete

Run again<br>Next attempt

Attempt 2

Image in<br>image labels out

The rule is unchanged<br>but the boxes now show the fruit image<br>instead of a string label.

Task: compare image with image<br>then choose the matching side.

Expected result: fewer conversions<br>faster reactions<br>fewer errors.

Prompt lesson: align the prompt format<br>with the data the model already has.

Fruit0 / 20

Correct0

Avg All0.00s

Avg Correct0.00s

left

right

Sort with visual labels

The only change<br>is the box label format.<br>Click a box<br>or use the keyboard arrows.

Start attempt 2

Attempt complete

Run again<br>Next attempt

Attempt 3

String labels<br>with warm-up

This attempt restores word labels.<br>Before the game starts<br>you get 30 seconds of examples:<br>fruit on top<br>text label below.

Task: use examples<br>to reduce the cost of unavoidable conversion.

Prompt lesson: few-shot examples<br>preload the mapping.

Compare this run with attempt 1<br>to feel the warm-up effect.

Fruit0 / 20

Correct0

Avg All0.00s

Avg Correct0.00s

left

right

Warm up first

A 30-second label-to-image pre-training screen<br>starts before sorting.

Begin 30s warm-up

Pre-training 30.0s

Warm-up complete

Play starts<br>when you choose to begin.<br>Click a box<br>or use the keyboard arrows.

Start play<br>Repeat warm-up

Attempt complete

Warm up and rerun<br>See techniques

What the game demonstrates

Prompting is like Coding

You are not only writing instructions.<br>You are designing the path<br>from input to answer.

Attempt 1

More transforms

The user sees a fruit image<br>converts it to a word<br>compares words<br>then acts.<br>More mental steps mean slower<br>and noisier execution.

to<br>apple

Attempt 2

Direct match

The label format<br>matches the input format.<br>The mapping is easier<br>so the same task consumes less attention.

to

Attempt 3

Few-shot warm-up

When conversion cannot be avoided<br>examples train the mapping<br>immediately before the task.

means<br>lemon

Techniques learned

Three prompting moves<br>that reduce<br>errors, latency, tokens, and cost

Minimize transforms

Ask in the representation closest to the input and output.<br>If the model has code, ask for code-shaped changes.<br>If it has a table, ask for a table-shaped answer.

Warm up with examples

When a transform is unavoidable<br>add a few compact examples.<br>Examples reduce ambiguity<br>better than long abstract rules.

Use analogies

Analogies describe the operation<br>instead of only describing the answer.<br>They let the model reuse<br>a familiar relation or pattern.

Remember ratio-proportion?

Boy

Girl

::<br>King

Queen

Why analogies work

Analogies give LLMs a direction.

In embedding space,<br>meaning is encoded as position.<br>A good analogy gives the model<br>a transformation to reuse.

Relationship: read the direction from boy to girl .

Apply: shift that same direction onto king .

Result: the shifted direction lands at queen .

boy

girl

king

queen

vector: boy -> girl

position: king

same vector, moved

boy -> girl<br>+ king<br>= queen

Apply it

Build prompts<br>that make the model do less<br>of the avoidable work

Toggle the techniques and watch the prompt change.<br>A longer prompt isn't the goal; The desired output is.

Patch review<br>Boot log triage<br>Field repair note<br>Weekly risk brief<br>Custom prompt

Minimize transforms Preserve source shape, names, and units.

Constraints Set scope, assumptions, and stop rules.

Output format Specify sections, order, and brevity.

Warm-up examples Show one compact mapping for unusual work.

Analogy Frame the relation only when it...

attempt image warm model fruit prompt

Related Articles