Skip to content

veryCoolTimo/imagegen-skills / skills/image-prompt

image-prompt

Flagged

Use when the user wants to turn a short idea into a rich, production-grade image-generation prompt — posters, landing-page or UI mockups, ads, editorial layouts, photoreal scenes, game screenshots, logos, or illustrations. Builds ONE structured, copy-paste-ready prompt optimized for OpenAI GPT Image 2.6 by default, following proven layered-composition patterns (named zones, hex palette, real font references, mood/finish, IP-safety). Triggers on requests like "сделай крутой промт для постера", "нужна картинка/постер/лендинг для…", "generate a prompt for a landing hero", "prompt for an ad / logo / game screenshot".

references/models/gpt-image.md

View on GitHub
# Model adapter — OpenAI GPT Image (default: gpt-image-2.6-sunburst)

How to render the 9-block content as a prompt tuned for OpenAI's GPT Image family, and how
to pick `model`, `size`, `quality`, and `complexity`. Distilled from OpenAI's official
"Image prompting" guide (GPT Image 2.6 edition, developers.openai.com/api/docs/guides/image-prompting).

Output is **prompt text only** (v1). Report the recommended params in the assumptions line so
the user can plug them into whatever tool they use.

---

## OpenAI GPT Image family (variants)

The prompting approach is the same across the family; models differ in speed, detail, params,
and lifecycle.

| Model | When to use |
|---|---|
| **gpt-image-2.6-sunburst** | **default** — base model: demanding detail, dense layout/text, precise edits |
| gpt-image-2.6-flare | smaller/faster model: drafts, volume, latency-sensitive work; try it once Sunburst passes |
| gpt-image-2.5-sunburst / 2.5-flare | previous generation; for existing 2.5 workflows (2.5 → same-named 2.6) |
| gpt-image-2 | existing validated workflows only (quality `low/medium/high/auto`; no `complexity`) |
| gpt-image-1.5 | **deprecated**, shuts down 2026-12-01; fixed sizes, `input_fidelity` low/high |
| gpt-image-1 | **retired** 2026-10-23 — don't target it; suggest 2.6 instead |

Pick by fit: user wants speed / many drafts → **Flare**; text-heavy, infographic, identity edit,
or "best possible" → **Sunburst**. If the user names "gpt-image-2" or just "openai", build for
2.6 Sunburst and say so (the prompt itself works on gpt-image-2 too).

## Parameters reference (2.6)

**quality:** `auto` | `low` | `medium` | `high` | `xhigh` | `max` — how much visual detail can be
rendered (texture, small lettering).
- `high` — dense/small text, infographics, slides with legends/axes/footnotes, close-up
  portraits, identity-sensitive work. **Default for text-heavy work.**
- `medium` — general default for photos/ads/illustration/logos.
- `low` — speed/volume/ideation.
- `xhigh` / `max` — only when `high` still misses a concrete detail requirement; not a
  guaranteed improvement, slower.

**complexity:** `low` | `medium` | `high` (default `high` when omitted) — how much the model
**plans** before rendering.
- Open-ended prompt (layout, text placement, props left to the model) → leave `high`.
- Fully specified prompt (layout + text + style already set — typical for this skill's 9-block
  output) → complexity makes little visible difference; leave default.
- Targeted single-detail edit → `low`, then check surroundings stayed intact.
- Tune quality and complexity **one at a time**: layout wrong → complexity; layout right but
  small text unreadable → quality.

**output_format:** `png` | `jpeg` | `webp`; generations also support **`svg`** (editable vector,
e.g. logos/icons; supports transparent bg; no streaming/compression) and **`gif`** (short loop;
`frame_count`, `frame_duration_ms`, e.g. 9 frames × 200ms at `512x512`).

**background:** `auto` | `opaque` | `transparent` (transparent needs PNG/WebP/SVG; check the
decoded alpha — a drawn checkerboard is not transparency).

**Edits:** omit `input_fidelity` (inputs are always high fidelity on 2.x). New `markup` param:
a transparent PNG overlay, same pixel size as the first input image, with circles/arrows
marking what to change (see `edit-remix.md`).

**size constraints:** `auto` or custom `WxH` — both edges multiples of 16, aspect between 1:3
and 3:1. (2.5 also documents: edge ≤ 3840px, 655,360–8,294,400 total pixels, above 2560×1440
experimental — treat as the safe envelope for 2.6 too and verify large sizes on your account.)

**aspect → size map** (use these defaults):

| Aspect / use | size | Note |
|---|---|---|
| Square | `1024x1024` | general-purpose default |
| Portrait / poster (2:3) | `1024x1536` | standard portrait |
| Landscape / hero (3:2) | `1536x1024` | standard landscape |
| Widescreen 16:9 | `1536x864` | decks, slides, game screenshots |
| Mobile 9:16 | `864x1536` | both ×16 ✓ |
| Premium wide 2K | `2560x1440` | reliability upper boundary |
| 4K UHD | `3840x2160` | experimental |
| Looping GIF | `512x512` | animated output |

If the user asks for an aspect not in the map, compute the nearest size where both edges
are multiples of 16 and the ratio is within 1:3–3:1.

---

## Prompting rules (baked in)

1. **Define the result.** Subject + intended use (ad / UI mock / infographic / poster) up
   front; composition, aspect, placement constraints. For complex requests: **scene →
   subject → details → constraints**, as short labeled sections, not one long paragraph.
2. **Maintainable format.** Paragraphs, labeled sections, JSON-like or tag lists all work —
   no magic syntax. Pick what's easiest to read and update (this skill: labeled sections).
3. **Visible details.** Concrete materials, shapes, textures, lighting, colors, and the
   visual medium. For wide/cinematic/low-light/rainy/neon scenes, specify scale, atmosphere,
   and color — not mood words alone.
4. **Photorealism:** include the literal **"photorealistic"** / "real photograph"; describe
   framing and texture (pores, wear, grain). Camera specs (35mm, 50mm lens) are appearance
   cues, not exact simulation. To avoid the glossy AI look: "honest and unposed, no
   glamorization, no heavy retouching" / "not an overly enhanced or cinematic movie-poster
   image".
5. **Composition:** framing, viewpoint, perspective, lighting, explicit placement ("logo
   top-right", "subject centered with negative space on the left").
6. **People:** body framing, relative scale, gaze, object interaction ("full body visible,
   feet included", "looking down at the open book", "hands naturally gripping the handlebars").
7. **Exact text:** literal copy in **quotes**; describe position + typography; spell
   unusual/brand words letter-by-letter; say **how many times** it appears ("render the
   tagline exactly once"); "no extra text". `quality: high` for small/dense text or multiple
   fonts.
8. **Artifact-spec framing** for slides/diagrams/charts/UI: name the exact deliverable,
   canvas, hierarchy; give the **real numbers and labels** in the prompt; for UI, describe the
   product **as if it already exists and shipped** (no concept-art language). Educational
   visuals: audience + lesson objective + required labels + what must NOT be included.
9. **Constraints block (always include):** "original, non-infringing design, no logos or
   trademarks, no watermark". For a brand-like mark: "generic / original, not resembling any
   real brand".
10. **Iterate deliberately:** one change per follow-up, restate what must be preserved; mention
    this to the user if they want to refine.

### Special outputs

- **Logo as SVG:** when the user wants a scalable/recolorable logo or icon, recommend
  `output_format=svg` (+ `background=transparent`). Prompt: "clean, vector-like shapes, strong
  silhouette, balanced negative space, flat, minimal strokes, no gradients unless essential;
  single centered logo with generous padding". For raster transparent: `png` + `transparent`,
  and add "clean alpha edges, no solid backdrop, scenery, checkerboard, or watermark".
- **Looping GIF:** give the subject a **starting pose it returns to**, describe one short
  motion arc, say whether the camera moves ("keep the camera still"), keep props anchored
  ("keep the lantern in its hands"), end with "make the motion loop smoothly". Params:
  `output_format=gif`, `size=512x512`, `frame_count`/`frame_duration_ms`.
- **Comics:** one concrete, action-focused visual beat per panel ("Panel 1: …").
- **Character consistency:** first prompt establishes the character (appearance, proportions,
  outfit, tone, plain background); follow-ups pass that image back with a
  `Character Consistency:` list and "do not redesign the character".
- **Historical/real-world context:** name place + exact date ("Bethel, New York, August 16,
  1969") and let the model infer period details; remind the user to check accuracy.

---

## Output template

Assemble labeled sections in this order. Omit sections that don't apply to the archetype
(e.g. TYPOGRAPHY for a candid photo; HUD only for game screenshots).

```
[Concept line — format + medium + fictional brand + reference aesthetic + intended use.]

BACKGROUND / CANVAS:
[aspect ratio, background treatment with hex, global texture]

SUBJECT:
[granular hero detail; IP-safe; real-texture cues if photoreal]

LAYOUT / COMPOSITION:
[named zones top→bottom, each with position + size% + tilt + content]

TYPOGRAPHY & TEXT:
[each text element: "LITERAL COPY" + named font + weight/tracking/case + hex + placement]

PALETTE:
[4-6 hex codes, each mapped to an element]

MOOD:
[adjective cluster + era/reference anchor]

FINISH:
[post-processing + authority/quality tag]

CONSTRAINTS:
Original design, no logos or trademarks, no watermark. Render each text string exactly once,
as written; no extra text. [+ any preserve/exclude notes]
```

For **photoreal scene / game screenshot**, swap LAYOUT/TYPOGRAPHY/PALETTE for the labeled
technical blocks: `SCENE:` `SUBJECT:` `ACTION:` `LIGHTING:` `CAMERA:` `HUD:` `TECHNICAL:`
(see `anatomy.md`, photoreal variant).

---

## Param selection heuristic

1. `model`: `gpt-image-2.6-sunburst` by default; `gpt-image-2.6-flare` if the user wants
   speed/drafts/volume.
2. Map the archetype's default aspect → `size` (table above); override if the user gave an
   aspect.
3. `quality`: `high` if the image has dense/small text, is an infographic/slide, a close-up
   portrait, or a poster with heavy typography; else `medium`; `low` only for drafts.
4. `complexity`: omit (default `high`) for generations; `low` for a targeted single-detail edit.
5. `output_format`: `svg` for logos/icons the user will scale or recolor; `gif` for loops;
   else `png`.
6. State them in the assumptions line, e.g.
   `model=gpt-image-2.6-sunburst · size=1024x1536 · quality=high`.

## Check the result (tell the user what to verify)

- Every word legible and matching the requested copy; diagram labels/relationships correct;
  numbers in slides are the ones you supplied (the model may add its own).
- Identities, product shapes, labels, reference details intact.
- An edit changed only what was requested; if a region must stay pixel-identical, composite
  the approved edit into the original.
- Transparency: the decoded file actually has transparent pixels.

## Sources
OpenAI "Image prompting" guide — GPT Image 2.6 / 2.5 / 2 / 1.5 / 1 tabs (developers.openai.com,
checked 2026-11-06). Verify against the live API before shipping: exact large-size limits on
2.6, GIF frame limits, and model lifecycle dates can change.
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