adapting-formats

Files (4)

• SKILL.md
• eval_cases.json · evals/
• ratios-and-safe-zones.md · references/
• recomposing.md · references/

SKILL.md (first lines)

# Resizing an ad

Take an image ad that already exists and produce it in the other shapes it has to run in.

**Each shape is regenerated, not cropped.** A crop cuts the headline in half and pushes the product out
of frame; the ad is rebuilt for the new proportions with the original passed in as a reference, so it
stays the same ad rather than becoming a different one.

## Workflow

### Step 1: Get the ad and the shapes

- **The ad** — the image file or URL, at the largest version the user has. Without it there is nothing
  to resize; ask for it and stop. Several ads at once is normal: each is read and planned separately,
  though they can all be generated in one call at the end.
- **The shapes** — the ratios, or the channel they run on. Where the user named a channel rather than
  ratios, take them from `references/ratios-and-safe-zones.md`. Where neither is given, ask which sizes
  they need, offering that table's channels and a free-text way out.
- **The product photo, where they have one** — a clean shot of the product on its own. It is worth
  asking for: the ad may show the product small, angled or partly behind copy, and the photo holds its
  identity far better than the ad can.

**Drop the shape the ad is already in.** An ad supplied at `1:1` and asked for the Meta family needs
`4:5` and `9:16` — regenerating its own `1:1` bills for a reconstruction of something the user already
has.

### Step 2: Read the ad

Run `image_analysis` on it once, and separate what the ad is made of — the parts a resize moves, not a
description of the picture:

- **The product** — what it is, and whatever has to stay identical wherever it appears: its shape,
  colour and label, the text printed on it, or anything else that would give it away as redrawn.
- **Every drawn string** — captured word for word, wherever it sits: a headline, a support line, a call
  to action, a price or a badge. These are regenerated, so a word read wrong here is a word wrong in the
  ad.
- **The background** — the surface, the scene, the light direction and the perspective, since this is
  what gets extended into any new area.
- **The look** — the palette, the character of the type, and whatever else makes it read as this
  brand's ad.
- **The layout** — what the eye lands on first, and how the parts sit relative to each other.

Take what this ad actually has rather than working down the list; an ad with no call to action doesn't
need one invented for the reading.

Read it once. A second analysis call bills again and tells you what you already have.

### Step 3: Plan each new shape

Read `references/recomposing.md` first — it owns this craft: what grows, what moves, what fills the
space that opens up, and the things that mark an ad as repurposed.

For each shape, decide and write down:

- **How far it is from the original.** A `1:1` into `4:5` is a nudge. A `1:1` into `9:16` is a real
  recomposition. A vertical into `16:9` or `21:9` inverts the frame and keeps very little — close to a
  new ad than a resize, and worth saying so before it is generated rather than after.
- **What fills the new area** — normally the background extended in keeping with its own light and
  perspective, plus deliberate negative space around the product. Never a new element invented to fill
  it.
- **Where each part goes**, including which part had to move out of the band the platform covers.
- **Whether the copy still fits.** The `9:16` safe area is roughly the middle half of the frame, so a
  copy-dense square ad often cannot move into it at a readable size. **Say that; don't shrink the type
  until it fits.** The answer is fewer words, and that is the user's call, not this skill's.

### Step 4: Show the plan and wait

Nothing has been generated yet, and everything after this costs money.

In one short message: the shapes being made, what fills the new space in each, anything significant
that had to move, and anything the shape can't carry — copy that won't fit, a ratio far enough from the
original that little survives. Then ask, and wait.

Where the user has said they don't want to be asked, say in one line what you're going with and carry
on.

### Step 5: Choose the model

**One model for every shape**, so the set comes back looking like one campaign rather than three ads.
A resize redraws every string the ad carries, so how well a model holds text is usually what decides it.

| Model | Use it when | How it wants the prompt |
| --- | --- | --- |
| `gpt-image-2` | the ad is a designed, text-forward layout — a headline, an offer, callouts | short labelled segments on separate lines, every string quoted exactly |
| `nano-banana-pro` | the ad leads on a photoreal product or a person and carries lighter text | connected sentences in one narrative paragraph, the drawn copy quoted |
| `seedream-5` | a photoreal ad another model refused or keeps failing | a description leading with whatever matters most |

Most ads carry real copy, so `gpt-image-2` is the usual choice. Where the user has named a model, use
that instead. Call `list_image_models` for what is actually available and for each model's
`reference_images` limit rather than assuming this table is the whole roster.

### Step 6: Write each shape's prompt

One prompt per shape, built from Step 2's reading and Step 3's plan, written in the shape the model
wants. Hold the product, the palette, the type and every drawn string identical across them — **only
the arrangement changes.**

- **State the new arrangement** — where each part sits in this shape, what the background extends into,
  and where the negative space falls. Say it as regions of the frame rather than leaving the model to
  infer it from the reference.
- **Keep the copy out of the platform's furniture.** Take the margins from
  `references/ratios-and-safe-zones.md` and state them. Margin cannot be added to a finished frame, so
  this is the one thing that cannot be corrected afterwards.
- **Quote every drawn string exactly**, in single quotes, with its size and position against the
  others.
- **Say the product is the one in the reference** and reaches the image unchanged — its shape, colour,
  label and printed text as they are.

### Step 7: Generate

Call `image_generate` with a `requests` list, **all of them in the same call** so they render at once.

- **One object per shape** — `prompt`, `model`, `aspect_ratio`, `resolution`, `reference_images`.
- **The original ad goes in `reference_images` every time**, and the clean product photo alongside it
  where the user supplied one. They carry the product, the palette and the type; the prompt alone will
  not.
- **`resolution` is `1k`** unless the user asked for a higher tier — raise it where the ad is dense with
  small type, which is where the extra pixels actually show.
- **The model from Step 5 on every object.** Take the accepted `aspect_ratio` values from

From SupercmoHQ/superCMO-skills, Apache-2.0.

More from SupercmoHQ/superCMO-skills