Σ 0:00▸ 0:00

Module 1 · Prompts as the Foundation of Computation · scripted

A Conversation Can Be Written Many Ways

by Jules White

20 minoutcomes: prompt-engineering
🎤 Talk view← Course

Overview

The last lesson showed one conversation programming one operation: a gap analysis. That conversation was written in ordinary English sentences. It did not have to be. The same task can be written in many different languages and grammars — plain prose, a labeled form, a structure that looks almost like code — and the form we choose changes how well the work comes out. This lesson looks at four conversations, each written differently, and then names what we are actually choosing when we write a prompt.

A conversation can argue against a plan

The gap analysis found what was missing. A different conversation asks for a different operation entirely — here, argue against the plan on the board:

Example 1 · The devil's advocate
user

Here is the plan our team put on the board this morning.

📷 the team's plan on the whiteboard

Argue against it. Take the strongest case you can that this plan fails, and give me the three most likely reasons.

assistant

1 — This is a list of ideas, not a plan. Every item names a thing to build. Nothing on the board says who does it, when it is due, or what it depends on.

2 — One item is circled, and nothing says why. If that is the first thing to do, the reason it beat the others is not written down — so the next person to look at the board will re-argue it.

3 — Several ideas assume data that may not exist yet. The board treats "use the data" as a step. Whether the data is there, and in what shape, is the risk — and it is not on the board.

✎ noteThe picture is the material. The instruction — "argue against it" — names the operation. Both are ordinary sentences.

A conversation can pull a structure out of a picture

Now a task with a different shape. We do not want an argument or an analysis; we want fields:

Example 2 · Reading a receipt
user

Read this receipt and return the date, the vendor, the total, and the currency.

📷a phone photo of a restaurant receipt

assistant

date: 2026-08-24 vendor: Hotel Alexandra total: 486.00 currency: NOK

✎ noteThe instruction lists the fields it wants, so the answer comes back as those fields. We did not ask for prose, and we did not get prose.

The same task, written as prose

Here is a third task — write a marketing plan — asked for the way most people ask:

Example 3 · The task in prose
user

We sell a waterproof duffel bag for people who commute by bike. Write me a persuasive marketing plan for it. Make it two or three paragraphs and really lean into the persuasion.

assistant

Every cyclist has had the same morning: the forecast said light rain, and the laptop got wet anyway…

The same task, written as a structure

Now the identical task, written in a different grammar — labeled inputs, labeled outputs, one setting per line:

Example 4 · The task as a structure
user
in:
  product: waterproof duffel bag
  buyer: people who commute by bike

out:
  marketing plan:
    persuasion: high
    length: 2-3 paragraphs
assistant

Every cyclist has had the same morning: the forecast said light rain, and the laptop got wet anyway…

✎ noteNothing here is code — no program runs it. It is still a conversation, and the LLM still reads it as words. It is the same request as Example 3, written in a different grammar.

Read Examples 3 and 4 next to each other. They ask for the same thing. The prose version buries the settings inside sentences: two or three paragraphs, really lean into the persuasion. The structured version puts each setting on its own line with a name. Neither is right in general. They are different languages for the same task, and they are easy to change in different ways — to make the second one longer, you edit one line.

Four conversations, four forms

Every example so far was a conversation. None of them looked alike.

The conversation Its form What it produced
gap analysis a document, then an instruction two named gaps
devil's advocate a picture, then an instruction three arguments
the receipt an instruction that lists fields four fields
the marketing plan prose, then the same thing as labeled settings the same plan, twice

The form is not decoration. It is part of the program.

We are choosing a language and a grammar

When we write a prompt we make two choices, and it helps to name them.

The two choices

languagethe words we use — the terms, the names, the level of detail
grammarthe shape we put them in — sentences, a list of fields, labeled settings, a form to fill in

We choose both, every time, whether or not we notice.

It is not wordsmithing

This is the part that is easy to get wrong. Working on a prompt is not hunting for magic words.

What we are actually doing

We are communicating a task to something that will act on it.

The goal is to state the task clearly and efficiently enough that the work comes back right more often.

Structure helps because it makes each part of the task visible:

  • The inputs are named, so nothing is assumed.
  • The settings sit on their own lines, so they are easy to see and easy to change.
  • The output is described, so we get what we asked for.

Prose helps when the task is short, or when the nuance lives in the sentences and a form would flatten it.

How we write the prompt affects accuracy, cost, and how much work it takes to change it later.

The idea to keep

A conversation is how we program computational thought.

The same task can be written in many languages and grammars.

The form we choose is part of the program — not the packaging.