# First Draft, Then Refine: Iterating and Judging AI Responses

> LOs: LO-S2-2 (Apply), LO-S2-3 (Evaluate)

## Introduction

Here is a habit that quietly costs people hours every week: they paste a prompt, glance at the response, sigh, and either ship it as-is or abandon AI altogether. Both reactions miss the point. The first response from any model is a **draft, not a deliverable** (see Slide 1). The real skill — the one that separates casual users from confident ones — is the small loop you run *after* that first answer: Prompt → Response → **Evaluate** → Refine. In this class you will learn how to judge a response in seconds and decide what to do next.

## Core Concept

Every time you read an AI response, you face exactly one decision with three branches. We call it the **Accept / Refine / Follow-up rule** (see Slide 2).

| Branch | When to use it | What to do |
|---|---|---|
| **Accept** | The output meets your bar — correct, on-format, usable. | Ship it. Stop the loop. |
| **Refine** | The answer is on-topic but misses the mark (wrong tone, wrong format, missing context). | Re-prompt with a targeted fix. Do **not** start over. |
| **Follow-up** | The answer is good and you now want to go deeper or branch out. | Ask the next question; keep the thread. |

When you choose **Refine**, resist the urge to rewrite everything. Apply at most one or two of these three targeted fixes:

1. **Add context** — who is the audience, what is the situation, what constraints apply.
2. **Add an example** — show the model one row, one sentence, or one snippet of the style you want.
3. **Tighten the format** — name the structure: "as a 5-bullet exec summary," "as a markdown table," "in under 80 words."

And critically: know your **stop condition** *before* you start. "Three bullets, board-ready, under 100 words" is a stop condition. "Make it better" is not. Without one, you will iterate forever.

## Worked Example

Imagine you are a BioTech CEO preparing a board update on a Phase 2 trial result (see Slide 3).

**Turn 1** — You prompt: *"Summarize our Phase 2 trial results."* The model returns a 400-word wall of text mixing endpoints, statistics, and dosing notes. It is technically correct, but no board member will read it. You **diagnose**: wrong audience, wrong format, wrong tone.

**Turn 2** — You refine with three precise changes:
- **+ audience:** "for a non-scientific board of directors"
- **+ format:** "as 3 bullets, then one risk callout"
- **+ tone:** "confident but not promotional"

The model returns a crisp, board-ready summary in seconds. Notice you did not rewrite the prompt from scratch — you *layered* three fixes onto the existing thread. That is the move.

Now try the practice case from Slide 4: you ask the model about mold flash defects and get a generic textbook list. Accept, Refine, or Follow-up? **Refine — and add context** (your specific resin, gate location, shot weight). Generic in, generic out.

## Common Pitfalls

- **Using Follow-up when you need Refine.** If the current answer is *wrong*, do not pile a new question on top of it — fix the current turn first. Otherwise the thread inherits the flaw.
- **Changing too many things at once.** If you add context, swap the format, *and* change the tone in one re-prompt and the result improves, you have learned nothing about what actually fixed it. Change one or two levers, not five.
- **Never stopping — perfectionism.** Without a written stop condition, every output looks like it could be 5% better. Set the bar before you start, and when the response clears it, ship.

## Recap

Iterating is not a sign that the model failed — it is the **method**. You now have a tool for **applying** prompts in a loop (LO-S2-2) and a rubric for **evaluating** each response with the Accept / Refine / Follow-up rule (LO-S2-3). Try the section quiz next; a couple of the questions are built around exactly this decision, and getting them right will lock the habit in.