6 words

A simple idea with a deep technological core: the user writes a question consisting of exactly six words, the platform generates an AI image as a cover, and the community must reply in exactly 6 words. Not 5, not 7 — exactly 6. The limit forces deep thought, and the results are surprisingly witty, wise, and poetic.

The platform operates within the Slovak-Czech community and as of August 2026 records 203 questions, 1,274 answers, and 39 registered users. Questions cover categories such as Relationships, Health, Personal Development, Culinary, Finances, or Self-Reflection — each of which is not only named but also visually portrayed by AI.

The biggest problem in generative graphics is the inaccuracy of direct prompting. An abstract question like "What is love?" produces generic outputs. I implemented a two-phase process:

1. Phase 1 — Prompt Enhancer: GPT-4 receives the question and a system instruction to deconstruct it into a precise visual descriptor — scene description, color palette, composition, mood.
2. Phase 2 — Image Generation: The optimized prompt enters DALL-E 3 at a resolution of 1792×1024px.

The results are significantly more consistent and visually striking than with direct prompting.

Every system event triggers a dedicated Messenger handler, ensuring lightning-fast UI responsiveness even during intensive AI operations:

• AiImageGenerateHandler — Prompt-on-Prompt + DALL-E 3 question cover generation
• AiAnswerGenerateHandler — GPT generates initial answers for every new question (seed content)
• AiCategoryGenerateHandler — GPT assigns the question to an existing or new category
• AiCategoryImageGenerateHandler — DALL-E 3 generates a cover for the new category
• AiNicknameGenerateHandler — GPT generates a pool of unique nicknames for new users
• SendDailyUserUpdatesHandler — daily email digest of new answers and likes
• SpecialDayNotificationHandler — special event emails with interactive buttons

Adding a question costs credits. Credits are replenished automatically via Symfony Scheduler with two recurring jobs:

• AddFreeCoinsProvider — replenishes credits according to rules every minute (cron: * * * * *)
• Daily Reports — cron at a random time between 8:00–20:00, dynamically generating a list of users with new activities

The scheduler runs without an external cron daemon — purely through Symfony Messenger workers.

Loupe — a typo-tolerant full-text search engine in pure PHP — indexes all questions. Minimum prefix length 3 characters, fuzzy matching, zero dependency on an external server.

AutoModerationService implements Weighted Report Score (WRS): each report carries weight based on severity (1–4), reporter trustworthiness (reporterTrust: 0.5–1.5), and time degradation. The system suggests an action: visible / soft_hidden / hard_hidden / quarantine. Reporter trust evolves based on their reporting history.

A custom Imagick-based HeroGeneratorService generates hero backgrounds for profiles and special pages. The algorithm:

1. Loads AI category images as tiles (600×400px)
2. Arranges them into a grid with a brick pattern (every second row offset by half a tile)
3. Applies a sigmoid gradient blend mask between tiles for smooth blending
4. Rotates the entire canvas by −15°
5. Crops the result to 2560×1440px, increases saturation, and exports as JPEG 80%

The result looks like a generative AI background, yet it is a deterministic composition of existing images.

The project is one of the first production deployments of the symfony/ai package. Individual AI agents are defined as Symfony services with the #[Autowire(service: ...)] attribute:

• ai.agent.image_prompt — Prompt Enhancer for DALL-E
• ai.agent.answers — seed answer generator
• ai.agent.category — question category classifier
• ai.agent.nickname — nickname generator
• ai.agent.optimize_categories — category optimizer

DALL-E 3 is called directly via PlatformInterface with a custom ImageUrl response format.