1

Break It Into Pieces

Tokenization

Before the model does anything else, it chops your prompt into pieces it can work with. Start by seeing your own prompt the way it does.

Paste it exactly as you typed it.

List the 3–5 words carrying the actual instruction. Everything else is filler the model has to work around.

Reflect If you deleted everything except those 3–5 words, would the output change much?
2

Find The Near Neighbors

Embedding

Similar words end up close together in the model's internal map. Test that against your own instincts before you assume it works the way you'd guess.

Not a synonym you looked up — your gut guess at what would land near it.

Reflect Swap the neighbor word into your prompt. Do you expect the output to shift a little or a lot? Why?
3

Label The Structure

Encoding

Encoding is just giving structure an address. Tag your prompt with CO-STAR so you can see what's actually there, and what isn't.

Click a tag for reference, then write which part of your prompt (if any) covers it.

Context
Objective
Style
Tone
Audience
Response
Reflect Which element is missing? Is that the part of your output that usually comes out weakest or vaguest?
4

Predict Before You Run It

Inference

The model is always predicting the most likely next step. Practice doing the same before you see its answer.

Rewrite it with the missing CO-STAR element added.

One sentence — what will the output contain? Write this before you run the prompt.

Reflect Where did it diverge? What's the one word or phrase you'd change next time to close that gap?
5

Debrief

Your Deconstruction, Reviewed

Here's everything you logged, pulled together in one place.

Your Prompt Deconstruction

Original prompt
Key words
Near neighbor
CO-STAR tags
Revised prompt
Prediction
What happened
Ongoing Practice

Pick one recurring prompt you use weekly — a status update, a client email draft. Log it below each time you use it, along with the one CO-STAR element you added and what changed.

Prompt CO-STAR element added What changed