Skip to content

gabelul/pixeltamer-gpt-image-skill

pixeltamer

Flagged

Generate, edit, and compose images with GPT Image 2.6, including transparent PNGs with a real alpha channel. Three backends — OpenAI API key (full feature set), codex CLI on your ChatGPT subscription (agent-driven, no key needed), or the direct ChatGPT image route on the same subscription token (--backend codex-images, PNG-only, opt-in). Four modes — one-shot generate, edit or inpaint an existing image, multi-reference composition (up to 16 inputs blended into one), and multi-image batch with verification. Use when the user asks to make an image, generate a poster, draw a mockup, design an icon or sticker, create an ad creative, build an infographic, blend reference images, edit or inpaint an existing image, or produce a cutout asset with a transparent background — "transparent PNG", "no background", "alpha channel", "works on any background". Not for: stripping the background off an existing image you want to keep (that's rembg or ImageMagick, not a generator); file operations that don't change the picture (compression, format conversion, resizing); reading, describing or OCRing an image; video.

references/prompting.md

View on GitHub
# Prompting gpt-image-2

GPT Image 2 (April 2026) is not the same animal as the previous generation of image models. The old "magic word" approach — stacking *4K, 8K, ultra detailed, masterpiece, trending on artstation* — is at best useless and often actively makes outputs worse. The new model rewards precise, instructional prompts written like a product spec, not like an incantation.

This doc teaches you how to write prompts that actually work for this model.

## What changed in GPT Image 2 (vs previous generations)

- **Strong instruction following.** Long prompts (50–300 words) don't lose detail anymore. Be specific.
- **Near-perfect text rendering**, including CJK and Arabic. Quote the exact string.
- **Custom resolutions**: any WxH where max side < 3840px and aspect ratio ≤ 3:1.
- **Edit endpoint does precise local edits**: tell it what to change AND what to preserve.
- **Multi-reference composition**: pass up to 16 reference images to `/images/edits` and the model blends them with consistent lighting and shadows.

## The canonical prompt structure

```
Intent / Use case  →  Scene / Background  →  Subject  →  Key details  →  Text content  →  Style language  →  Constraints
```

Worked example:

> *Create one **pitch-deck slide** titled "Q3 Revenue Performance" that looks like a real Series A board-meeting slide. Layout (landscape 16:9): title top-left in 36pt Inter dark gray, two-column body with a 60% bar chart on the left and three KPI cards on the right, footer with a logo placeholder bottom-right and page number "03". Chart: clean vertical bars for Q1–Q3 2026, y-axis $0–$8M, three bars at $3.2M, $4.8M, $6.5M, muted blue palette, subtle gridlines, no shadows. KPI cards show: "+34% YoY", "189 new accounts", "$42K ACV". Style: white background, Inter typography, tight spacing, no decorative elements, no clip art, no gradients, no stock photography.*

Notice what's NOT in there: no "highly detailed", no "8K", no "masterpiece", no "trending on artstation". Every word is doing concrete work.

## The seven habits of high-quality prompts

### 1. State the intent first

Open with what the image is FOR, not what's IN it. The model has different "modes" for different intents and your opening word selects the mode:

- *Create one pitch-deck slide titled…*
- *Create a realistic SaaS dashboard UI mockup for…*
- *Create a children's book illustration of…*
- *Create a product hero photograph of…*
- *Create a character reference sheet for a 2D platformer…*
- *Create a vertical infographic titled…*
- *Edit the input image: change ONLY [X]…*

This single change moves quality more than any style word you could add.

### 2. Quote the exact text

For every character that should appear in the image, quote it. Specify font style, weight, size, color, and position:

- ✅ `Headline (top center, 64pt bold sans-serif, dark gray): "Maker Era"`
- ❌ `a poster about Maker Era`

For non-Latin scripts, name the script flavor explicitly and add `no extra characters anywhere except the quoted text` to suppress decorative-looking nonsense.

### 3. Use photographic / spec language, not praise language

Praise language ("stunning, beautiful, ultra-detailed, professional, premium") tells the model nothing concrete and pushes outputs toward stock-photo-ad aesthetics. Spec language tells the model where to put things and what they look like.

| Replace | With |
|---------|------|
| "beautiful lighting" | "soft north-window light from the left, gentle wraparound fill" |
| "professional" | "Inter 14pt, 8px grid spacing, single accent color #6366F1" |
| "high quality" | "shot on 50mm f/2.8, fine 35mm grain, neutral white balance" |
| "detailed" | "visible ceramic glaze imperfections, fine wood grain on the counter" |
| "modern" | "flat vector, monoline 2px stroke, generous negative space, off-white background" |

### 4. Specify "change ONLY X / keep everything else" for edits

The edit endpoint is the killer feature, but only if you tell the model what to preserve. Without explicit preservation, lighting drifts, perspectives shift, faces change.

Template:

```
Edit the input image: replace ONLY [thing] with [new thing of same shape/proportions].

Preserve exactly:
- Camera angle and framing
- Room lighting and shadow direction
- Floor reflections and contact shadows
- All other furniture, walls, and decor
- Color temperature and white balance

Render: [new thing details].

Do not: add new objects, change saturation, restyle the room,
or alter any text or signage in the scene.
```

For people, add a `Preserve exactly: facial features, identity, hair, expression` block.

### 5. Pick the right resolution for the use case

| Use case | Size | Notes |
|----------|------|-------|
| PPT / web hero / YouTube thumbnail | `1536x864` (true 16:9) or `1792x1024` | Title-safe area on one side |
| Print poster, mobile cover, story | `1024x1536` (2:3) or `1024x1792` | Vertical typography |
| Square (icon, IG post, logo) | `1024x1024` | Sprites/icons that need transparency: generate with `--background transparent` (API backend) — see `references/transparency.md`. Chroma-key is the fallback, not the plan |
| High-res square / print poster | `2048x2048` | Slower; only when needed |
| 4K landscape / wallpaper | `3840x2160` | Top of supported size range |
| 4K portrait / vertical poster | `2160x3840` | Same |

`2K` and `4K` are accepted as shorthand for `2048x2048` and `3840x2160`.

Four constraints on a custom `WxH`, all enforced locally before the request goes out:

| Rule | Value |
|---|---|
| Both edges divisible by | `16` |
| Longest edge | ≤ `3840` |
| Total pixels | ≤ `8,294,400` |
| Aspect ratio | ≤ `3:1` either way |

That pixel cap is exactly `3840x2160`, which is also why `2880x2880` is the practical square ceiling — it lands on the same number.

Wrong aspect ratio = wasted generation. Decide BEFORE you write the prompt.

### 6. Iterate, don't stack

If the first generation is 80% there, **change ONE dimension** and re-generate. Don't bolt three new clauses onto the prompt — you won't know which one helped.

| If the result is… | Change THIS only |
|-------------------|------------------|
| Wrong vibe | Style anchor (named artist / movement / film) |
| Too generic | Composition (specific camera, angle, framing) |
| Lifeless | Lighting (direction + color temperature) |
| Cluttered | Negative space + remove adjectives |
| Off-brand color | Add hex codes or named palette |
| Text wrong | Re-quote the text and add `no extra characters` |

For exploration, fire 3–4 variants in parallel (`-n 4`) — cheaper than serial single calls.

### 7. Default to high quality, drop only when iterating fast

`high` is the production default. Use `low` / `medium` for speed-over-fidelity rough comps where you're iterating layout. Use `auto` if the API host charges differently per tier and you don't know what you need yet.

## Style vocabulary (for when you DO want a style anchor)

A named reference beats five vague adjectives. Anchor to one of:

### Photography
*Annie Leibovitz (editorial), Steve McCurry (documentary), Wes Anderson (symmetric pastel), Gregory Crewdson (cinematic suburban), Henri Cartier-Bresson (decisive moment B&W), Tim Walker (theatrical fashion).*
*Film stocks*: Kodak Portra 400, CineStill 800T, Ilford HP5, Fuji Velvia.
*Lens / camera*: 50mm f/2.8, 85mm f/1.4, Hasselblad medium format, 16mm fisheye.

### Illustration
*Flat vector (à la Stripe / Linear marketing), isometric line art, children's book watercolor (Beatrix Potter), Art Nouveau (Mucha), Art Deco poster (Cassandre), Bauhaus geometric, Saul Bass mid-century, risograph two-color print, gouache spot illustration.*

### Anime / Asian art
*Studio Ghibli watercolor, Makoto Shinkai cinematic anime, Mamoru Hosoda warm tones, Pixiv illustration, traditional ukiyo-e woodblock, Chinese gongbi line painting, ink wash.*

### 3D / digital
*Blender Cycles claymation, Octane render, Unreal Engine 5 cinematic, low-poly PS1-era, voxel art, glassmorphism UI render.*

### Game / retro
*16-bit SNES pixel art, NES 8-bit limited palette, Game Boy 4-shade green, PICO-8 16-color, Stardew Valley pixel style, Hades vector painted, Hollow Knight hand-drawn.*

### Painting
*Oil painting in the style of John Singer Sargent, impressionist plein air (Monet), expressionist (Egon Schiele), baroque chiaroscuro (Caravaggio), Hopper americana, Rothko color field.*

**Rule of thumb**: if you can't name the style precisely, name the *reference* — an artist, a film, a movement, a magazine, a game.

## Composition vocabulary

| Need | Words |
|------|-------|
| Distance | extreme close-up, close-up, medium shot, full-body, wide shot, establishing shot |
| Angle | eye-level, low angle, high angle / bird's-eye, dutch tilt, over-the-shoulder |
| Framing | centered, rule of thirds, asymmetric, symmetrical, copy area on the [side] |
| 3D POV | front view, 3/4 view, isometric, top-down, side-profile, axonometric |
| Layout | generous negative space, tight crop, full bleed, vignette |
| Lens | shallow depth of field, fisheye distortion, tilt-shift, telephoto compression |

For posters and slides, **always** state where the negative space goes ("copy area on the right third for headline + subtitle").

## Lighting vocabulary

| Direction | Words |
|-----------|-------|
| Natural | golden hour, blue hour, harsh noon, overcast diffused, moonlit, north-window light |
| Studio | softbox key + rim light, three-point lighting, beauty dish, butterfly lighting |
| Mood | chiaroscuro, low-key (mostly shadow), high-key (mostly highlight), backlit silhouette |
| Color temp | warm tungsten, cool fluorescent, magenta neon, sodium street lamp, candlelight |
| Effects | volumetric god rays, lens flare, atmospheric haze, hard cast shadows |

Pick a single light direction with a single color temperature. Two light sources = okay, three = mush.

## JSON vs prose — when each wins

A common question because the freestylefly / EvoLink galleries are full of JSON-style prompts. The honest answer (cross-referenced from OpenAI's cookbook, fal.ai's guide, and the skill's own controlled A/B test):

**JSON does NOT beat well-structured prose for the same content.** OpenAI's cookbook explicitly says "minimal prompts, descriptive paragraphs, JSON-like structures, instruction-style prompts all work, as long as the intent and constraints are clear." The dev.to JSON-advocacy article that gets cited everywhere provides zero side-by-side comparisons.

The skill's own A/B test (documented in `prompt-patterns.md` § 1) found: **on the codex backend, structurally different prompts with the same content produced byte-identical PNGs.** Codex's reasoning loop normalizes the prompt before calling gpt-image-2, so JSON's structural advantages collapse on that path. On the API backend the structural value still applies — but only as a *human/agent* organization tool, not because the model "prefers" JSON.

**Decision rule:**

| Use JSON when | Use prose when |
|---|---|
| 6+ visual subsystems (luxury product photo, brand identity system, complex scene reference board) | Single-image creative work (mascot, illustration, photo, poster) |
| Batch generation with templated slot-filling | One-off generation |
| Agent composing prompts programmatically | Human writing creatively |
| Brand consistency across many touchpoints | Personality + gesture flow narratively |

For mascots, characters, and most one-off creative work: **stick with structured prose** (labeled sections, exact quoted text, hex colors). For commercial spec systems and multi-touchpoint brand work: escalate to `prompt-patterns.md` § 1 for the JSON schema.

The legacy magic words ("8K UHD", "ultra-detailed", "极致锐度") that appear in many viral JSON prompts in the repos are noise either way — JSON doesn't fix bad content.

## Front-load the first 50 words (gpt-image-2-specific)

Per OpenAI's official cookbook and fal.ai's prompting guide, gpt-image-2 weights the first ~50 words of a prompt disproportionately. The model treats whatever comes first as the anchor and everything else as secondary detail. This means:

- **Lead with the SUBJECT description**, not meta-framing like "this is a mascot for an iOS app."
- **Don't open with intent labels** ("INTENT: Brand mascot…"). Open with a 2–3 sentence concrete description of what's in the image.
- **Atmosphere, mood, and style anchors come AFTER subject**, not before. The model needs to know *what* before *how*.
- **Use case** (ad / mockup / cover / icon) belongs near the start but as one phrase, not a full block. Example: *"App icon mascot. A small designed [character description]…"*

This applies to every prompt the skill produces, not just mascots. If your prompt opens with three lines of meta-framing before the subject appears, rewrite it.

## Replace praise language with observable spec language

Per OpenAI's cookbook: words like *cute, charming, beautiful, premium, stunning, professional, iconic* are praise without observable content. They tell the model nothing concrete and pull outputs toward generic-stock-image defaults. Replace with phrases that name a *testable* visual property.

| Don't say | Say instead |
|---|---|
| "cute character" | "kind expression, gentle eyes" |
| "charming mascot" | "brave but warm demeanor, head tilted 5° camera-left" |
| "beautiful design" | "[material]: brushed paper, flat folds, two-tone shading" |
| "stunning poster" | "asymmetric type block, generous negative space, single accent color" |
| "professional look" | "Inter Bold typography, 8px grid, single accent #1F6B6B" |
| "playful and fun" | "head tilt, weight on one leg, gaze meeting viewer" |
| "premium quality" | "two shading tones per surface, no gradients, crisp edges" |
| "iconic" | "recognizable in solid black silhouette at 64×64px" |
| "high-end commercial" | "shot on Hasselblad medium format, soft north-window light, neutral white balance" |

The test: if you can't write a unit test for the adjective ("contains exactly two shading tones"), it's praise — replace it.

## Picking the right anchor for your context

Style vocabulary above tells you *what to write*. This section tells you *what to reach for*. The same word — "premium," "flat," "playful" — pulls toward different aesthetics depending on context. A wrong-context anchor produces a technically-correct but useless image (e.g., a 3D vinyl figure when you needed a flat app icon mascot).

| Where the image will live | Reach for | Avoid |
|---|---|---|
| Mobile app icon / mascot | Flat geometric vector — Reddit Snoo, Duolingo Duo, MetaMask origami fox, GitHub Octocat, Things 3 | 3D vinyl, designer toy, photoreal |
| SaaS marketing site illustration | Flat editorial — Linear's character set, Notion brand illos, Stripe brand illustrations | chibi, anime, over-cute |
| Premium B2B / fintech | Minimal geometric — Stripe, Things 3, Linear, Arc Browser | colorful playful animal-cartoon |
| Magazine / book cover | Era-anchored editorial — 1990s Wallpaper, early Wired, Aperture, New Yorker | generic stock-cover, ad-look |
| Product photo / e-commerce hero | Studio commercial — name a real campaign era (Aesop, Glossier, Apple product page) | painterly, illustrated |
| Luxury / fashion / fragrance | Named houses — Tom Ford, Dior, Aesop, Hermès campaign | bright pop-art, kawaii |
| Children's app / illustration | Soft watercolor — Beatrix Potter, Sago Mini, Khan Kids | corporate-flat, photoreal |
| Designer collectible / merch | Matte vinyl — Bearbrick, Funko, Medicom, Pop Mart, Daniel Arsham | flat-vector (looks cheap) |
| Game IP character | Match game's art direction explicitly (pixel/low-poly/painted/anime) | generic 3D blob |
| Editorial mascot for content brand | Restrained character — Mailchimp Freddie, MailerLite, MetaMask fox, Duolingo Lily | cute-overload sparkles rosy-cheeks |
| Loading screen / single-color mark | Monoline silhouette — Octocat solid, Twitter old bird, Things 3 mark | painterly multi-tone |
| Pitch deck / SaaS slide | Flat, restrained — Apple Keynote-grade, Stripe slide deck | clip-art, stock-photo |

**Rule**: name ONE specific anchor in the prompt, not a list. Specificity is the contract; "Linear-inspired flat-vector friendly modern" is mush. The model has seen Snoo; it has not seen "playful flat character."

For mascot/character work specifically, see `recipes/mascot.md` — it expands this matrix with the failure modes the model defaults to in each context.

## Color & palette

State a palette explicitly:

- Named: *muted earth tones, monochrome blue, pastel mint and coral, jewel tones, sepia*
- Specific hex / Pantone: *cream #f5e6d3 background, deep navy #1a2b4c accent, single accent #6366F1*
- By reference: *Wes Anderson palette, 80s Memphis, Bauhaus primary red/yellow/blue*

For brand work, ask for hex codes before generating. Saves regen cycles.

## Negatives — say what NOT to include

GPT Image 2 obeys explicit negatives. Use them when the model keeps adding things you don't want:

- *no text, no watermarks, no logos*
- *no humans, no faces, no hands*
- *no extra fingers, no warped anatomy*
- *no busy background, plain backdrop*
- *no decorative gradients, no stock photography aesthetic*

What you should NOT use as negatives:
- ~~lowres, bad anatomy, text, error~~ — these are old-model magic words; they're noise to GPT Image 2.

### Negatives — what a strong list actually looks like

The vague negatives above work for one-offs. For brand and commercial work, write **8–15 named failure modes**, each targeting an actual bad-default behavior gpt-image-2 keeps producing in your category. "No busy background" is a wish; "no decorative side panels, no rounded-corner cards, no thick borders, no large background graphics" is a constraint.

See `prompt-patterns.md` § 3 for worked negative blocks per style (Apple-keynote, graffiti sketch, premium typography poster, brand touchpoint system).

## Prompt template (copy / adapt)

```
Create a [INTENT — slide / poster / icon / illustration / mockup / photograph]
of [SUBJECT].

Layout: [where things go, sizes, alignment, negative space].

Text: [exact quoted strings, font style, size, color, position;
       OR: "no text".]

Style: [single named reference; OR: medium + era + adjectives].

Color: [palette by name or hex codes].

Lighting: [single direction + color temperature].

Constraints: [explicit negatives].
```

## Three habits that mark a pro prompt

1. **Open with intent**, not subject ("Create a pitch-deck slide…" beats "A chart and some KPIs").
2. **Quote every character** that should appear in the image, exactly once.
3. **Specify what to preserve** alongside what to change in every edit prompt.

## When to escalate beyond this file

Reach for `prompt-patterns.md` when the request hits any of: building the prompt programmatically (slot-by-slot rather than free prose), a complex image with many independent subsystems, one of the doctrine categories (UI / infographic / brand identity / e-commerce hero / architectural render / scientific atlas / conceptual typography poster), or brand consistency across many images. Patterns covered there: JSON-config schema, role-based opener, specific-negative blocks, phase-based structure, auto-deduce closure, signature integration.

## Anti-patterns to drop

| Don't write | Why |
|-------------|-----|
| "8K, 4K, ultra detailed, masterpiece" | Magic words from older models; ignored or harmful here |
| "trending on artstation" | Same |
| "professional, beautiful, premium, stunning" | Praise language with no instructional content |
| Long lists of stacked styles | "Watercolor + 3D + cyberpunk + manga" produces mush |
| Negative prompts like "lowres, error, bad anatomy" | Pre-GPT-Image-2 noise |
| Re-describing the whole image in an edit | The model already sees it; you're just shifting things |
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall