← Back to all articles

9/22/2025

From Sourcing to Spotlight: How We Curate This AI Prompt Gallery

AI Image Prompt Team#curation#workflow#quality

Our gallery is often mistaken for a passive mirror of social media trends. In reality, every entry is the product of a deliberate, auditable curation pipeline. We start upstream by scouting creators across X, Xiaohongshu, and Discord communities who repeatedly demonstrate technical command of Nano Banana or GPT-4o. We maintain a contributor ledger that tracks authorship, licensing signals, and communication notes so we can verify attribution quickly when updates are needed.

Once a promising case surfaces, we request the original prompt text, any parameter overrides, and raw reference assets. Screenshots alone are never enough; we need the underlying process to confirm the example is reproducible. If reference files are proprietary or time-limited, we redact sensitive data and substitute anonymised placeholders in the JSON dataset while preserving the structural intent. Transparency beats mystery hacks every time.

Data hygiene and translation

Curation is a bilingual effort. The Nano Banana repository ships with seven README translations, so we cross-check terminology to keep English and Chinese entries aligned. When we load a case into our dataset generator, we strip emoji noise, convert smart quotes to ASCII, and normalise spacing. We also assign canonical slugs that include the model name to avoid collisions between repositories. This frictionless naming scheme is what allows the website to generate sitemap entries, canonical URLs without .html suffixes, and analytics tags automatically.

Every text asset passes through our consistency glossary. For example, "reference image" and "source A" must be used in the same way across cases, or users will misinterpret the instructions. When creators provide multilingual notes, we store both versions and mark which was authored first; that helps translators prioritise nuance rather than performing literal word swaps.

Visual verification lab

After textual cleanup comes visual verification. We reproduce each prompt using the declared model build. For Nano Banana edits, we pay special attention to identity retention across sequences—if twenty hairstyles share the same subject, we render them in batches to ensure the facial geometry remains locked. For GPT-4o graphics, we test the prompt at multiple resolutions to observe where composition drifts or text rendering fails. Any systemic quirks are documented in the notes field so readers know when they may need to adjust seeds or sampling steps.

We also convert outputs to WebP and run lossless compression benchmarks. Although the PRD allows JPEG, we strive to publish assets that load instantly on mobile. Lazy loading is enabled on the website, but high-quality compression reduces dependence on network speed.

Metadata and SEO alignment

A curated gallery is only useful if people can find what they need. Every case is tagged with style and theme metadata derived from the prompt vocabulary. Those tags feed the filter UI on the homepage and the sitemap generator. We craft alt text following the "AI image prompt – Title (input/output)" convention so screen readers communicate context while keeping search engines happy. Canonical URLs omit .html to align with our SEO guidelines, and every page description starts with "#1" to satisfy the specification our stakeholders set for this project.

We monitor performance metrics like LCP and CLS to ensure the waterfall layout remains snappy. Larger inputs are preloaded when necessary, and we stagger animations to avoid jank. The result is a cyberpunk aesthetic that still respects accessibility.

Ethics and update cadence

Curation is a living responsibility. We respect takedown requests, credit corrections, and new licensing information promptly. If a creator releases updated prompt notes, we regenerate the JSON and redeploy through our Cloudflare pipeline. Every case page lists its author handle and links back to the original post so traffic flows to the source community. By treating curatorship as both archival work and platform stewardship, we keep the gallery trustworthy for practitioners who rely on it for production inspiration.