Dali by Lulu

Score your creative against what's actually winning in the ad market — before you spend the credit.

Most AI generation failures are predictable. A weak prompt, an off-formula creative — you can't tell until after you've burned the token. Dali scores it first, and it doesn't grade against opinions or generic "prompt tips." It grades against a real, living corpus of proven-winning ads — creatives still running in the market months after launch, scraped, embedded, and ranked. Two jobs:

• score_prompt — judge the prompt before you generate (craft: camera, motion, lighting, model-native language).

• score_creative — judge the actual image against proven winners (does it look like what converts, and what's missing).

Every wasted generation has a real cost — a Seedance retry is ~$6. The live dashboard tracks what the community has saved by catching bad creatives before they burned a credit.

You: "make a video ad for our glass serum bottle"

dali::score_prompt(prompt, "veo3")
→ 8/100  Grade: F
→ no camera move · no motion · no lighting · 8 words
→ Verdict: Generic stock footage guaranteed. Enhance first.
→ enhancement_brief included (score < 70):

  ① lead with camera — Veo 3's #1 lever: "Slow dolly", "Orbital push"
  ② describe physics: "a drop falls", "liquid ripples", "glass refracts"
  ③ lighting type + quality: "warm backlight", "rim-lit edges"
  ↳ [Camera]. [Subject + motion]. [Lighting]. [Mood]. [No text.]

✦ YOUR LLM rewrites using the brief:

  "Slow orbital push around a glass serum bottle on white marble. A single
   amber drop falls in extreme slow motion, catching warm backlight. Macro:
   liquid gold ripples outward from impact. Rim-lit edges, soft studio
   diffusion. Premium, clinical. No text."

dali::score_prompt(enhanced, "veo3")
→ 91/100  Grade: A  ✓ Safe to generate.

The real winning data layer

This is what makes Dali more than a prompt linter. The scores are grounded in real ads that are actually winning, not hand-written rules.

How the corpus is built — longevity is the outcome signal. We scrape the public Meta Ad Library. An ad still running months after it launched is one the advertiser keeps paying for — a proven winner. That "still-running-after-N-days" longevity is a market-validated label you can't fake, and it's the spine of the whole dataset.

What's in it, today:

The pipeline (offline → serving). The tools never scrape or embed on the fly — they read pre-built stores:

scrape Meta Ad Library         → proven winners (longevity label)
      → Gemini vision           → creative attributes (lighting, format, before/after, offer…)
      → prevalence SQL          → winning-pattern lift per vertical (winners vs baseline)
      → Vertex embeddings       → BigQuery VECTOR_SEARCH (nearest proven winners, cosine)
      → graph edges (Memgraph)  → (:Pattern)-[:WINS_IN {lift, n}]->(:Category)

So when score_creative runs, it embeds your image and finds the actual winning ads it most resembles by full visual signature — then tells you which winning attributes you're missing. When enhance_prompt runs with a category, the rewrite brief is backed by real market lift ("before/after shows up in 78% of winning wellness ads, 4× baseline"), not craft opinion.

Honest scope. The winner label is longevity (a strong market-validated proxy), not per-ad conversion rate — measured CVR validation is in progress. The corpus grows on a schedule, so coverage per vertical keeps deepening. What you get today: your creative scored against what's demonstrably surviving in the live market.

dali::score_creative(image_url, "beauty")
→ score 62/100  — partial resemblance to proven winners
→ looks_like:      Frøya Organics (ran 411d), tashportcosmetics (884d), Face Reality (346d)
→ what_to_change:  winners use "before/after" 4× more · offer-visible 1.8× more
→ defects:         none
→ Verdict: Partial — strong resemblance, but add the high-lift attributes before spending.

Related MCP server: Prompt Auto-Optimizer MCP

Contents

• The real winning data layer

• Install

• Tools

• Supported models

• Platform supersets

• Why model-specific?

• MCP resources

• Contributing

Install

Hosted MCP — connect once, scores every prompt and creative:

# Claude Code
claude mcp add --transport http dali https://dali.getlulu.dev/mcp

// Cursor / Windsurf — .cursor/mcp.json or windsurf settings
{
  "mcpServers": {
    "dali": { "url": "https://dali.getlulu.dev/mcp" }
  }
}

// stdio-only clients — npx wrapper around the hosted server, no Python needed
{
  "mcpServers": {
    "dali": { "command": "npx", "args": ["-y", "dali-mcp"] }
  }
}

→ Full install guide with all clients

Self-hosted — local, no auth required:

pip install dali-mcp
claude mcp add dali -- python -m dali.server

The self-hosted package exposes the prompt-scoring tools locally. The creative-scoring tools (score_creative, analyze_winning_formula) and the winning-ad corpus run on the hosted server — connect via the hosted MCP to use them.

Tools

Score the creative — against real winners

Score the prompt — before you generate

The graph brain & meta

Supported models

Video

Sora 2 (OpenAI): API shutdown September 24, 2026. Do not build new dependencies on it — use Runway or Kling instead.

Image

Imagen 4 (Google): deprecated — use gemini-3.5-flash with image output. Dali still scores legacy Imagen prompts via the imagen model key but don't build new things on it.

Platform supersets

Higgsfield and Runway are aggregator platforms — they proxy multiple underlying models under one API. The model you pick matters more than the platform name:

Dali scores for the underlying model's native prompt language, not the platform wrapper. Pass the model name (veo3, kling, seedance…), not the platform name.

Why model-specific?

Generic prompt optimizers don't know that:

• Veo 3.1 needs camera movement specified above everything else

• Kling 3 supports multi-shot scene labels natively in the prompt

• Flux responds to camera body and lens names like a photographer ("Sony A7 IV, 85mm f/1.4")

• Midjourney V8.1 reads prose + parameters, not keyword lists

• Higgsfield simulates physics — you describe materials in motion, not motion abstractly

• Minimax uses [Pan left] bracket syntax for camera moves — plain text camera commands are ignored

• Ideogram V4 needs text quoted exactly in the prompt for typography accuracy

• Wan 2.7 generates native audio — include sound descriptions alongside visuals

Dali has a separate scoring rubric and rewrite brief for each model. Your LLM does the creative rewriting — Dali provides the intelligence.

MCP resources

creative://guide/veo3       → Veo 3.1 camera language guide
creative://guide/seedance   → Seedance UGC motion guide
creative://guide/kling      → Kling multi-shot + expression guide
creative://guide/runway     → Runway motion-first guide
creative://guide/wan        → Wan 2.7 audio + motion guide
creative://guide/minimax    → Minimax bracket camera guide
creative://guide/higgsfield → Higgsfield physics-motion guide
creative://guide/sora       → Sora 2 guide (API shutdown Sep 24, 2026)
creative://guide/flux       → Flux photography brief guide
creative://guide/midjourney → Midjourney V8.1 + parameters guide
creative://guide/ideogram   → Ideogram V4 typography guide
creative://guide/firefly    → Firefly 5 commercial content guide
creative://guide/imagen     → Imagen 4 guide (deprecated Aug 17, 2026)
creative://models           → All models overview

Contributing

Model guides live in dali/data/guides/{model}.json on the hosted server. Found practitioner patterns that consistently produce high-grade results? Open an issue with the model, the pattern, and a sample prompt + result. The best contributions come from Reddit, Discord, and YouTube — real practitioners, not official docs.

→ Prompt best practices by model — cheat sheets, do/don't tables, top patterns per model → Dali creative flow skill — install this skill so your LLM follows the score → enhance → generate workflow automatically

MIT License · Built by Lulu · dali.getlulu.dev