I Gave GPT-6 Astra 600,000 Real Design References

Technological concepts & product features (what the video tests)

• Core problem tested: Even with a “best” coding/research model (GPT-6 Astra), AI-generated marketing sites often look generic—repeating common layouts/styles/SaaS design patterns.
• Hypothesis: A stronger model alone isn’t enough; better results come from giving the model real design context via reference research.

• Workflow design: A two-step AI web-design pipeline:

  1. Draft-only phase (no visual references): GPT-6 Astra generates an initial landing page from a text prompt/brief.
  2. Research + redesign phase (with 600k visual references): GPT-6 Astra connects to an MCP server backed by Mobbin to research UI patterns, then produces a revised landing page using those findings.

• References source: Mobbin library described as 600,000+ UI screens, accessed through a Mobbin MCP server.

• Model & tooling references:

  • Built inside Codex with the Mobbin plugin/MCP.
  • Optional image generation mentioned: “images 2.0 image generation model” via Codex.
  • An “Annotate” feature is used to iteratively apply specific edits to the generated design.

---

Video’s guide/tutorial-like setup (how to reproduce)

Mobbin MCP setup steps (high-level)

1. Create a Mobbin account → go to Settings → MCP tab.
2. Copy a one-line URL from Mobbin’s MCP settings.
3. In Codex: open plugins, search for the Mobbin plugin, add it, sign in and authorize.
4. Confirm Codex is connected to Mobbin’s 600k design references.

Research phase prompt behavior

Instruct the agent to use Mobbin MCP to research design patterns relevant to the current Agent Hub landing page, and require a report first, including:

• Where the current page feels generic
• Recurring patterns from relevant examples
• Recommended visual direction
• Concrete changes and “why”

Redesign phase

After the report, instruct Codex/GPT-6 to completely redesign the landing page:

• Overhaul visual style
• Match brand/category expectations
• Use reference-driven changes

---

Reviews/analysis of results (what changed and why)

Step 1 outcome (no references)

The first landing page is clearly AI-generated and resembles default templates (even likened to Bootstrap-style patterns). While it is responsive and includes the requested sections and a product UI preview, it still fails the “unique, believable category presentation” test—sticking to generic B2B SaaS conventions.

Step 2 outcome (Mobbin research + redesign)

The redesigned page becomes more personalized and category-appropriate, with:

• More intentional page structure
• More specific visual direction
• Updated/adjusted product presentation
• Improved interaction details (e.g., hover effects and a more tailored hero/product section)

The agent’s research report flags concrete issues like:

• Hero section too general
• Card sections missing key details
• Claims lacking proof

The agent cites relevant reference interfaces from products/tools such as Clay, Cursor, Vercel, and LangChain. Users can also click into references inside Mobbin to inspect full pages and similar patterns.

---

Iteration features demonstrated (manual refinement after redesign)

Using Codex annotate to request targeted UI changes, including:

• Turn a “trust” row into a rotating carousel
• Pause carousel on hover
• Add play/pause controls
• Replace placeholder text elements with real company logos in chips
• Remove an awkward/unnecessary bottom section

Result: The final page looks more editorial/personable and closer to an established company marketing site.

---

Key takeaway / conclusion from the analysis

• Main finding: A stronger model (GPT-6 Astra) speeds up generation, but it doesn’t automatically improve “taste.”
• Most reliable workflow: Don’t go “brief → finished site” immediately. Instead:
  • Research first (with real visual references)
  • Identify patterns that matter
  • Choose what fits the project
  • Then build/redesign

The video frames this as a more solid and reliable AI web design workflow going forward.

---

Main speakers or sources

• Speaker: The video’s primary creator/reviewer (unnamed in subtitles) running tests with GPT-6 Astra and Codex.
• Sources/tools referenced: OpenAI GPT-6 Astra, Codex, Mobbin (via Mobbin MCP), and design references from Clay, Cursor, Vercel, LangChain.