Image Generation Prompt Engineering — Universal

Generic prompt-engineering skill for any modern image generation / editing model. Works model-agnostic by default; switches to model-specific guidance when the user names a model.

Strong prompts share the same anatomy across models — subject, environment, lighting, camera/lens, style, format, and (for edits) what to preserve. What differs is syntax surface: Midjourney uses --flags, GPT Image 2 likes structured JSON, GPT Image 2.5 rewards role-assigned references and lock lists, Nano Banana parses natural-language addressing of reference images, SDXL leans on weighted token blocks plus negative prompts, FLUX rewards prose density, Ideogram is the strongest at typography.

When to Use This Skill

• User asks to write or improve an image prompt for any model
• User wants to edit an uploaded image (outfit swap, background change, face-preserving restyle, object addition/removal, inpainting)
• User requests identity-preserved portraits, celebrity integration, aging sequences, multi-character group shots
• User wants a bento/infographic layout, isometric diorama, exploded-view product poster, magazine cover, comic panel, storyboard grid, e-commerce main image, YouTube thumbnail
• User needs style-specific prompts: cinematic, Kodak Portra, anime, 3D chibi, pixel art, Baroque frame, ukiyo-e, Swiss design, cyberpunk, RAW iPhone, 90s point-and-shoot, 35mm direct flash
• User mentions "preserve face", "same face 100%", "change only the background", "keep pose", "multi-view", "action figure box", "exploded view"
• User wants Raycast-friendly prompts with {argument name="..." default="..."} placeholders
• User explicitly names a model (Midjourney, MJ, SDXL, FLUX, GPT Image 2 / 2.5, Nano Banana, etc.)

Model Routing

Whenever the user names a model in their query, read the matching file under references/models/ for syntax surface and quirks, then layer it on top of the universal anatomy below.

Always start with references/models/_index.md if you need to scan all routes at once.

Two Capability Tiers (general)

Per-model mapping of "Standard" vs "Pro" is documented in each references/models/<name>.md. Examples:

• Nano Banana = Standard; Nano Banana Pro / Nano Banana 2 = Pro
• GPT Image 1 / DALL·E 3 = Standard; GPT Image 1.5 = Standard+; GPT Image 2 = Pro; GPT Image 2.5 Flare = Pro, Sunburst = Pro+ (precision edits, dense exact copy)
• Midjourney v6 = Standard; v7 = Pro
• SD 1.5 = Standard; SDXL / SD3 / SD3.5 = Pro
• FLUX.1 [schnell] = Standard; FLUX.1 [dev] / [pro] = Pro
• Seedream 4.0 / 4.5 = Standard; Seedream 5.0 = Pro

Rule of thumb: default to Pro tier for infographics, posters with headlines, multi-character scenes, structured JSON/YAML prompts, and CJK text. Standard tier suffices for single-subject edits and short prompts.

Universal Prompt Anatomy

Every strong prompt stacks these blocks. Order is flexible; labels are optional when context is clear.

[Subject]        who/what — reference handling ("keep face 100%")
[Action/Pose]    what they're doing, body orientation
[Wardrobe]       clothes, accessories, materials, color
[Environment]    location, props, ground, background wall
[Lighting]       key/fill/rim, time-of-day, quality (hard/soft), color temp
[Camera/Lens]    85mm f/1.4, drone top-down, orthographic, eye-level
[Style ref]      film stock (Kodak Portra 400), director, era (2000s), render engine
[Format]         aspect ratio (1:1, 16:9, 9:16), resolution (1080x1080, 8K)
[Preserve]       (edit mode) what must NOT change
[Negative]       anti-artifacts: "no text", "no second object", "no warping"

Minimum viable template — generation

A {style/medium} of {subject} {action} in {environment}.
{Wardrobe + props}.
Lighting: {key + rim + mood}. Camera: {lens, angle}.
{Aspect ratio}, {resolution}. {Style ref}.

Minimum viable template — image edit

{Edit verb} the {target region} of the attached image:
{specific change}.
Preserve: face/identity, pose, outfit, background lighting — zero alteration.
Only modify: {explicit region}.
Match the original {lighting logic / color grade / perspective}.

Reference Inputs (per model)

How to address an uploaded reference image varies. Universal pattern works across most natural-language models; flag-based models need explicit syntax.

For exact syntax see the matching file under references/models/.

Category Catalog

Detailed templates and examples live in references/.

• references/patterns.md — universal vocabulary: lighting, camera, ratios, style tags, composition idioms, single-object rules, multi-reference fusion, negative-prompt vocabulary
• references/templates.md — fill-in templates for 18+ common categories (portrait, product, poster, diorama, infographic, comic grid, exploded-view, e-commerce hero, YouTube thumbnail, character sheet, storyboard, etc.)
• references/identity-preservation.md — face-lock phrases, failure modes, fixes, multi-person handling, mixed-style face preservation, aging sequences
• references/structured-prompts.md — JSON / YAML / XML formats and which models parse them reliably
• references/text-rendering.md — putting readable headlines, quotes, CJK in the image, per-model tier limits
• references/editing-workflow.md — non-destructive editing, inpainting, multi-image fusion, before/after, character consistency across panels
• references/examples.md — curated working prompts per category, attributed, with the model that produced each
• references/workflows.md — Direct Generation, Content Illustration (article/video → image), and Remix/Personalization conversation patterns
• references/external-corpora.md — pointers to public prompt libraries (YouMind 12k+, EvoLink ~3k, Awesome lists) when the user wants proven prompts rather than fresh composition
• references/models/ — per-model quirks, syntax surface, parameters, anti-patterns (gpt-image-2-5.md adds Flare/Sunburst choice, API size/quality rules, 10 patterns, 12 templates, known failures, review checklist)

Quick Rules (model-agnostic)

1. Paragraph + constraint lines beats tag soup for natural-language models (Nano Banana, GPT Image 2 / 2.5, FLUX, Imagen). For Midjourney and SDXL, dense comma-separated tag blocks plus flags still work.
2. Specify aspect ratio explicitly — 1080x1080, 16:9, 9:16, or --ar 9:16 for Midjourney. Don't rely on defaults.
3. Identity anchor every edit. "Same face 100% from reference", "keep facial features exactly consistent", "do not change the face".
4. Non-destructive edits need an allow-list. Name what changes AND what is preserved — most models otherwise drift on wardrobe/pose.
5. Lens + lighting = photorealism. 85mm f/1.4, Kodak Portra 400, three-point lighting, golden hour rim light, harsh on-camera flash all carry real weight on every model.
6. Single-object constraints must be repeated. If you want exactly one object, write an explicit strict-single-object block (see references/patterns.md#single-object-rule).
7. Text in image = quote it exactly. Wrap exact strings in double quotes: serif font reading "Stay Hungry, Stay Foolish". For CJK, prefer Pro-tier models.
8. Use [PLACEHOLDERS] (square brackets, uppercase) for reusable templates. Use {argument name="x" default="y"} for Raycast snippets — supported as a literal placeholder by Nano Banana, GPT Image 2 / 2.5, and any natural-language model.
9. For bento/infographic layouts, enumerate modules (M1…M8) with content type per cell — strongly favors Pro-tier models.
10. Negatives go at the end as a NEGATIVE: list or Do NOT: bullet — except SDXL/Midjourney where negatives have dedicated syntax (--no for MJ, dedicated negative prompt input for SD).
11. For natural / un-AI looks: explicitly request "natural skin texture, flyaway hairs, slight asymmetry, no glamour retouching, no beauty filter, no overly polished AI aesthetic". GPT Image 2 / 2.5 and Nano Banana Pro both respect these.
12. One change per edit turn + a lock list. Name the single change, enumerate what stays, allow only the physically necessary side effects (contact shadows, reflections). Branch variants from a saved master. Preservation is visual, not pixel-identical — composite when pixels must match.
13. Review every output against a concrete check — text at full and phone size, object/people counts, identity before styling, product geometry, edit scope, decoded alpha.
14. For cross-image consistency (same character across N panels): describe the character once in detail, then reference back as "the same character from panel 1" — works on GPT Image 2 / 2.5 and Nano Banana Pro; weaker on others.

Anti-Patterns

• Tag-salad prompts on natural-language models (portrait, 8k, masterpiece, beautiful, highly detailed) — under-perform vs structured paragraph
• Asking for "photorealistic" without lens/lighting — produces generic render
• Editing without a preservation clause — model mutates face/pose
• Multiple conflicting aspect ratios in one prompt — one wins, unpredictably
• Face-swap with only facial description (no ref image) — low fidelity
• Packing > 10 named subjects in one prompt — identity drift past ~5 on Standard tier, past ~8 on Pro
• Embedding exact text > 15 words on Standard-tier models — spelling breaks
• Mixing Midjourney --flags into a Nano Banana / GPT Image / FLUX prompt — flags get parsed as literal text and pollute output
• "Polished AI" defaults when shooting for candid lifestyle — explicitly veto glamour retouching
• Expecting an image model to export a GIF, ZIP, vector or HTML — generate the sheet, then split/assemble with code (see models/gpt-image-2-5.md Pattern 6)
• Trusting "no text / ONLY this text / no watermark" without checking — negatives are honored only part of the time
• Forcing JSON on a tiny prompt — {"subject":"red apple"} is worse than A red apple on a white table, studio lighting, 1:1.

Conversation Modes

Four distinct entry points handled in references/workflows.md:

• Direct Generation — user describes an image; you compose a prompt from anatomy + template + per-model file.
• Content Illustration — user pastes article / video script / podcast notes; you analyze theme + tone, match to a template, and move into Remix.
• Inspiration → Prompt (reconstruction) — user shares an image they like; you write a new brief from what is visible and list the unknowns (references, earlier turns, edits) instead of pretending to recover the original prompt.
• Remix / Personalization — after a template is chosen, ask only the relevant personalization questions (gender, age, setting, mood, profession) and remix the template with the user's content.

If the user just wants proven, image-attested prompts rather than a freshly-composed one, point them at the searchable corpora in references/external-corpora.md (12,000+ prompts at YouMind, ~3,000 GPT Image 2 cases at EvoLink, ~250 GPT Image 2.5 cases across wangrunlin / LaplaceYoung / VulcanEon, plus Awesome-list browsable indexes).

Workflow

1. Ask user — generate from scratch, edit an existing image, illustrate content, reconstruct a prompt from an inspiration image, or browse proven prompts? Which model? Which tier? Aspect ratio?
2. If a model is named, read references/models/<name>.md first.
3. Identify category → pick from references/templates.md.
4. If editing with a reference image: lock identity + add preservation clause first (see references/identity-preservation.md).
5. Fill subject → environment → lighting → camera → format → style.
6. Add negatives only for known failure modes (not prophylactically). Use the model's native negative-prompt mechanism if it has one.
7. For text in image, follow references/text-rendering.md and downgrade to a Pro-tier model if the string is long or non-Latin.
8. Output the prompt in a single fenced block; user pastes into the target model UI/API. For Midjourney, end with the flag block; for SD/FLUX, output positive + negative prompt separately. For GPT Image 2.5 on the API, list model (Flare or Sunburst), size, quality, background as parameters outside the prompt.
9. Close with a one-line Check: — the first thing the user should inspect in the result (spelling, count, identity, geometry, edit scope, alpha).