SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsimage-prompt

image-prompt

Design & creation

View repository

Updated 10-15 19:47

About this skill

🐾 Gongnyang Prompt Kit VOL.2 — gpt-image-2 prompt compiler

Compile ambiguous requests into $imagegenfinished Korean production prompts. For generation/mass production use [codex-imagegen](for 1 image, codex directly).

Workflow

  1. Infer and fill in missing decisions (category, cut type, subject, style, composition, text, AR). The only things to ask about are the original Korean copy text, brand names, and sensitive material.
  2. Determine which file to read from the routing table; formats are in §Format A/B.
  3. Write while following the 9 ironclad rules, and at the end AR token.
  4. For high-value deliverables, before responding run node scripts/check_prompt.mjs validation (ok:true).
  5. When linked to generation, pass only the compiled prompt (raw original text prohibited).

Expansion examples: poster→C3, pictorial→C1 B, key art→C11, icon→C9 ("no text"), product→C4, comic→C10, PPT→C12 (16:9).

Output contract: single=body + trailing AR x:yonly (without explanation) · multiple= Title/Category(Cn)/Cut type/Prompt per entry · generation request=compile silently then call the tool.

Routing table (sole routing point)

Request signal Category/Format File to read
Solo figure pictorial·editorial C1·Format B references/editorial-hwabo.md (lookbook·sequence·21 fashion types +references/style-taxonomy.md)
Typo poster·text is the image itself TP1~TP17 references/typo-poster-router.md→references/typo-poster/TPn-*.md 1개
Type specimen·glyph set·hand-cut wordmark·print gradient metal TP15~TP17 references/typo-poster-router.md→ corresponding TPn-*.md 1개
Promotional merchandise·brand poster·"well-designed poster" P1~P12 references/promo-router.md→references/promo/Pn-*.md 1. Card-news density grammar prohibited (aesthetic death)
Cover plate (album cover·book cover·packaging label)·color-field division·painterly cover·signboard-type collage P9~P12 references/promo-router.md→ corresponding Pn-*.md 1개
Poster·key art·infographic·card news·comic·field guide·icon·beauty·campaign·mockup C2~C11 references/category-patterns.md Corresponding §. C6·C7=density default·breakthrough tactics §C6
Presentation·slide deck C12 references/category-patterns.md §C12
Mood ("make it look good"·"luxury"·"cinematic") 룩 L1~L9 references/look-presets.md Drop in 1 preset
Draft diversification·mass-production concepts·"differentiation"·"start from the concept" M/R/X/T축 references/concept-axes.md Variation on 1 axis
Text placement·font·grid·dense text — references/typography-layout.md
Camera·lighting·color vocabulary — references/photo-vocab.md
jsonl batch·model facts·finished examples·codex skeleton·8-section variants — references/jsonl-and-examples.md

Routers (P/TP) select 1 pattern and then load only one corresponding file. If multiple rows match at the same time, upper rows take priority(table order = priority) — for boundary cases, the boundary/intersection reference section of each router takes precedence.

Ironclad rules

  1. Leading [AR x:y SIZE wxh] bracket prohibited. size is only via the API parameter (jsonl size), and only one at the end of the prompt AR x:y . Slot tokens ([PERSONA_LOCK] type) are authoring-only — any residue = disqualification (E-SLOT-LEAK).

  2. Scene exclusions are all positive-form — gpt-image-2 actually renders scene negatives (crowd→"one person, alone", background→"clean solid-color background"). The only exceptions are the two lanes; not a workaround but compliance steering.

    Tier Condition Allowed phrasing
    Tier-0 default Always all-positive, 0 negations
    Tier-1 text guard Only when rendered text is present 7-item whitelist: no extra words · no duplicate text · no invented glyphs · no watermark · no logo · no extra text · verbatim, no extra characters
    Tier-2 editorial lane Only when explicitly declared (no heuristic promotion) SAFETY_ASSERT (positive-form, subject clause) + NEGATIVE_TAIL (once, immediately before AR) pair structure

    Tier-1 combination formula (the only emission form): All text appears once, perfectly legible — no duplicate text, no extra words, no invented glyphs, no watermark. Tier-2 fixed strings and pair rules (order-preserving subset, tail standalone prohibited) canonical source = references/editorial-hwabo.md §3 (not included here). Negative: Labels are prohibited in all tiers (E-NEG-SECTION).

    What to remove Lane
    Scene elements (people, objects, background, props) Positive-form restatement (Tier-0)
    Text render defects (duplication, ghosting, watermark) Tier-1 whitelist
    Policy safety assertion (editorial) Tier-2 pair

  3. SD-era deprecated vocabulary prohibited. masterpiece/best quality/8k/4k/uhd/trending on artstation/ultra-detailed/highly detailed/sharp focus, weights (word:1.3), --ar/--v, body §, empty adjectives (beautifully/emotionally/luxuriously/refinedly/beautiful/stunning). Stage specification ("award-level/like a professional/top-tier") is equivalent; the standard lies outside the prompt. Reduce to numbers (margin %·60/30/10·hierarchy singular)·body reactions·concrete examples (concept-axes.md §dead-word reduction).

  4. Reduce equipment specs→results. Instead of EXIF·equipment names, use "shallow DoF, background falls off softly"·"warm key + cool rim". (Fashion Lens character:·Director signature:is an exception as result+anchor.)

  5. Numbers are locked in. HEX palette (3–5 colors per cut), Kelvin, key:fill 1:2 ratio.

  6. 1 row = 1 cut = 1 call. One-canvas grid/matrix prohibited; multiple cuts use N rows.

  7. Ideal skin prohibited → "natural skin texture, visible pores, subtle film grain".

  8. Real trademarks·person references prohibited, use fictional brands/personas.

  9. Absolutely no post-processing of text after generation. Text is rendered inside the image via the prompt (quoted copy + role labels + free authoring zone). Compositing code onto PNG (PIL·ImageMagick·SVG/HTML·canvas) is strictly prohibited; font·kerning·tone will look off. For text errors, revise the prompt and regenerate (typography concretization→2048x2048+quality high→copy reduction, in that order).

Master template (Format A, 6 sections)

Pure description + HEX. Core visual information at the very top (first section = strongest attention). Header # 1. Scene OK, body § forbidden.

# Section What / Length
1 Scene Who, what, where, what. Core first. 60–120 words
2 Camera Viewpoint, distance, lens character (result description). 15–30 words
3 Lighting Direction, soft/hard, shadows, rim light (no equipment names). 10–25 words
4 Color grading Palette + color temperature + HEX 3~5개. 10~20어
5 Texture/Medium Medium, texture, surface response, post-processing. 10–20 words
6 Text-in-image (optional) "따옴표 카피" + font, size, position + legibility once. 0–25 words
— Trailing At the end AR x:y token only

Format A / Format B

  • Format A — 6 labeled sections(table above): overall structures (posters, key art, infographics, compendiums, card news, comics, etc.).
  • Format B — pictorial flat comma-style short sentences: no labels, one comma paragraph, 350–450 characters, default AR 2:3(1024x1536). 12 slot types order, Tier-2 → editorial-hwabo.md.
  • Routing: solo person pictorial/glamour editorial → B, everything else → A.

Size lock (6 codex runner types)

API is custom, codex ($imagegen) path only 6 types safe. auto forbidden, unified within chapter.

ar size ar size
1:1 1024x1024 16:9 1792x1024
2:3 / 3:4 / 4:5 1024x1536 9:16 1024x1792
3:2 / 4:3 1536x1024 Dense/multi-cut 2048x2048

2:3only exact ratio, 3:4·4:5is vertical approximation. Unsupported values (1024x1280 etc.) also to the nearest 6 types.

API hard constraints, transparent background (1.5 fallback): jsonl-and-examples.md §1.

Text rendering key points

  • Copy is fixed in quotes + font, size, position, HEX, one language per line (no mixing Korean + English). 2+ copies separate roll labels (headline/subhead/callout), specify position, size, font per roll.
  • Rendered text gets Tier-1 combination formula once (iron rule #2), same copy written twice = rendered twice, forbidden.
  • Korean spell-out forbidden (hyphens render as letters). Accuracy lever = canvas size (2048 measured 12/12). Only English coinages ALL CAPS as last resort. Labels contain only actual wording.
  • Dense text (3+ copy blocks/small letters) is quality: high+ large side pairing (1536/1792,2048x2048).
  • Korean/English mixed: Korean = narrative, mood, cultural nouns, Korean copy, English = technique tokens, tier phrases, hybrid = Korean skeleton + English tokens. Details photo-vocab.md §8.

Validator

Before response/generation node scripts/check_prompt.mjs <file>(stdin possible).

  • --tier <0|1|2> enforced / --jsonl record / --api E-SIZE-LOCK→warning (hard constraints maintained) / --test self-test.
  • Tier priority: --tier > jsonl tier > lane("editorial"→2) > heuristic (rendered text→1, otherwise 0). Tier-2 cannot be promoted by heuristic.
  • Output {ok, format, tier, errors[{code,msg,hint?}], warnings[{code,msg}]}, errors 0=exit 0.
  • Errors: E-NEG-TIER(undeclared higher tier),E-SLOT-LEAK·E-SIZE-LOCK, negatives, leading brackets, SD vocabulary, weights, slash flags, missing trailing AR.