Y2K streetwear drop poster

For the LLM writing the next image prompt. When the brief says "drop poster", "streetwear poster", "poster-as-landing-page", "merch-drop image", "hype graphic", "album-cover-style still", or anything that needs a bold character-driven brand poster with one huge readable word — the rules below are the register. This guideline is the prompt reference; the [/poster] skill is the end-to-end workflow.

When to apply

Single still images that announce a brand or a drop: a giant exclamation wordmark, one chest-up character, sticker-collage accents, a vivid gradient. Works for any aesthetic on the same skeleton — Y2K skate-surf is the validated default, but vaporwave / grunge-zine / arcade-90s / clean-DTC all ride it.

Do not apply to: photoreal human portraits with no poster framing (use photoreal-studio-portraits), product-only renders (cgi-product-renders), or anything that should read as a quiet editorial photo rather than loud key-art.

Model picks

Goal	Model	Why
Final poster (the wordmark + small copy must read crisp)	openai/gpt-5.4-image-2	Premium typography: the beveled wordmark, the footer strip, and the 4-line body copy all stay sharp. Caps at 1 concurrent — serialize batches.
Fast palette / composition exploration	google/gemini-3-pro-image-preview	Cheaper, ~half the latency, tolerates 4+ concurrent — but smudges small embedded text, so finalize the chosen variant on gpt-image.

Aspect ratio: pass --size 1080x1920 for a vertical landing poster or --size 1080x1080 for a square social post. pica generate forwards this as the selected provider image_config.aspect_ratio, which gpt-image honors — the in-prompt hint alone is ignored (it would default to 1024²). The square cut is the strongest feed format for this layout; the vertical cut is the landing cut.

The fixed architecture (do not restructure — only fill the slots)

Every poster in this register keeps this top-to-bottom stack:

1. Top banner — small tracked uppercase intro line in the accent color.
2. Wordmark — a MASSIVE chunky slab-serif word with a thick 3D bevel shadow and a small chrome highlight per letter, tightly tracked, slightly tilted with imperfect hand-drawn terminals. One short shoutable token (BOOM / DANG / YIKES / WOW / OOPS).
3. Hero photo (~60% of height) — cinematic chest-up portrait, gaze slightly up, casual-cool. An original hand-scrawled blob glyph on the chest (lowercase brand word inside it), and a product held at mouth level (lollipop / popsicle / soda can / mic). Soft rim light matching the background, warm skin tones, film grain.
4. Sticker collage — exactly 3 vinyl stickers, each with a thin white border + soft cast shadow + tape corner, scattered at playful angles: one star with a product macro, one with the blob face, one small star with a pixel "NEW DROP"-type label. More than 3 and the eye has nowhere to rest.
5. Copy block — condensed uppercase headline + 4 lines of italic-serif body copy in cream + a filled CTA pill.
6. Footer strip — very thin, full-width, tracked uppercase: BRAND · EST. YYYY · <process tagline> · WWW.<domain>.

Finish layer, always on: subtle film grain + slight vignette + faint CRT warmth + vintage screen-print feel.

Swap slots (the only things that change between posters)

• brand / tagline / domain — the fictional or user-owned label.
• exclamation — the wordmark word.
• hero — person + wardrobe + expression (vary age / gender / styling).
• product — the prop at mouth level. Keep it generic/unbranded unless the user owns the brand (otherwise the reference-required gate fires).
• palette — background gradient (2 hex) + accent (wordmark/headline) + bevel-shadow color + cream body + charcoal linework + product color. Always state explicit hex.
• blob glyph — an ORIGINAL friendly creature (smiley blob, cyclops cloud, lopsided star). Always write "original, NOT any existing brand" or the model drifts toward a trademarked smiley mark.
• aesthetic — the register (Y2K streetwear default; others ride the same scaffold).

Failure modes

• Smudged wordmark or body copy → you finalized on gemini; redo the keeper on gpt-5.4-image-2.
• 403 "Key limit exceeded" → parallel gpt-image calls; serialize (cap 1).
• Glyph looks like a real brand → you dropped the "original, NOT any existing brand" clause.
• Word unreadable → the wordmark accent does not contrast the gradient hard enough.
• Apostrophes break the shell → write the prompt to a .txt and pass --prompt "$(cat file.txt)" instead of inline single quotes.

The three examples here (BOOM / DANG / YIKES) are the same scaffold with three hard palette / hero / product swaps — read one alongside the source prompt in workspace/projects/loud-kids-poster-001/ before drafting a new one.