# 0:35 · Iterate Like a Director

_LOFTai › Your First Hour_

**Goal:** turn one lucky result into a deliberate one — and then teach Loft what you learned so you never have to say it again. Ten minutes. This is the lesson that separates using Loft from using a prompt box.

## Three levers, in order of leverage

**1 · The words**

Cheapest and most direct. Change one thing at a time — light, or framing, or treatment — so you learn what actually moved the result.

**2 · The inputs**

Show, don't describe. A reference image pins down style or subject far better than another paragraph of adjectives.

**3 · The model**

When a prompt keeps almost-working, the model may simply not do that. Go back to the gallery rather than fighting it.

## Step 1 — Regenerate with one change

1.  Expand your completed queue card and click **Regenerate** — it brings the whole setup back, prompt and parameters intact.
2.  Change _one_ thing. Swap "dusk" for "hard noon light", or "wide establishing" for "low-angle close".
3.  Generate again, and compare the two results side by side in the asset's Files.

Do this twice. Two deliberate revisions teach you more about a model than twenty random rerolls.

## Step 2 — Bring in a reference

Now use the image you marked **Reference** in lesson 2. Pick a model that takes image input (an image-to-image or image-to-video type pill). Its setup form shows **input slots above the prompt**, each labeled by the _role_ it plays — First Frame, End Frame, Reference Images, Input Video, Reference Audio.

1.  Click a slot to open the picker; it browses **this asset's own files**. Click **Pinned** to see what's pinned in the Assistant instead. Select your reference and click **Done**.
2.  Each tile carries a **status badge**. By default the picker offers only **Published** files, plus **Selected** ones when they belong to this asset. The status dropdown widens that if you need a New or Dropped take.
3.  Notice the **@-mention badge** it gets — `@Image1`, `@Image2`, and so on.
4.  Name the badge in your prompt to say what each image is _for_: _"match the color palette of @Image1"_. Attaching an image without saying its job is the most common way references get ignored.

Note: Some models offer **scenario modes** — "First / Last Frame" vs "References" — as a toggle above the inputs. Only the active scenario's slots are submitted, so you can't accidentally mix incompatible inputs.

## Step 3 — Keep the language that worked

When a phrasing earns its keep, click **Save current prompt as primer** above the prompt box. A **primer** is a reusable prompt block you inject on demand — yours (My Primers), the project's, or system-wide. Next time you generate, it's one click instead of retyping.

## Step 4 — Write your project playbook

This is the highest-value five minutes in the whole hour. Go to **Project Detail** and click the **book icon** in the header to open the **Project Playbook**.

A playbook is free-form standing guidance the Assistant reads _every turn_ — house style, default sizes, naming conventions, model preferences. Write three or four short declarative lines, like:

-   House look is 35mm film, warm, slightly hazy. Avoid clean digital renders.
-   Default to 16:9 for shots, square for character reference.
-   Always generate one image first; only make variations when I ask.
-   Name shots SHxxx, three digits, zero-padded.

Keep it short and declarative — it's read on every request, so concise rules beat long prose. Playbooks layer: project, then asset, then your personal defaults, each refining the last. Primers are pull; playbooks are always-on.

Tip: The tell that a playbook is working: you stop retyping the same clause in every prompt. Anything you've typed three times belongs in a playbook, not a prompt.

## Step 5 — Try it in words instead

Now open the **Assistant** from the header and just ask: _"generate another take of this shot, tighter framing, same light."_Same generation machinery, no form. The Assistant reads your playbooks on every turn, so the guidance you just wrote is already in play.

One phrasing rule worth internalizing: **imperative phrasing executes, desiderative phrasing discusses.** "Generate X" acts. "I'd like to explore X" talks it through first. Choose deliberately.

## Checkpoint

-   Your asset holds three or more takes, and you can say why each differs from the last.
-   At least one generation used a reference image, named in the prompt by its @-badge.
-   Your project playbook has a few lines in it.
-   You've asked the Assistant for something and gotten a result.

Note: Reference topics for this lesson: **Generation**, **Primers**, **Playbooks**, and **Assistant**.