Lab · How I prototype with AI2026

Prototypes built on real code.

Before I design a screen, Claude reads the backend. Then I prompt Claude Design over many rounds per module, fine-tune the details from a business point of view, audit every build against the code, and put a working prototype in front of stakeholders in about a week.

  1. 01 Inputs Requirements, legacy designs, backend code
  2. 02 Read the code Claude Code
  3. 03 Prompt & refine Many rounds per module
  4. 04 Build Claude Design
  5. 05 Audit Against the code
  6. 06 Demo Stakeholders and client
Running underneathEvery decision, prompt and open question goes into a shared Obsidian vault, the team's brain.
  • 8modules in one interactive prototype
  • 2backend codebases read before designing
  • 1wkper enterprise prototype
  • 1module designed straight from the code, with no written spec
01 — The system

Six steps, every time

Shown with Ratings & Reviews v2, an enterprise module for UpStart Commerce.

  1. 01

    Start from requirements and the real backend

    Not from screens alone. Ratings & Reviews started from a written requirements doc, the legacy Figma designs and the two backend services behind it.

    DOCWritten requirementsRatings & Reviews v2
    FIGLegacy designsFigma
    SVCBackend service 1Reviews and ratings
    SVCBackend service 2Moderation
  2. 02

    Let Claude read the code

    Claude Code reads the codebases directly and pulls out the GraphQL schemas, domain models and events. That way every screen is grounded in what the API really supports.

    Claude Code
    > read both backend services and list what the API supports
    ✓ GraphQL schemas
    ✓ Domain models
    ✓ Events
    → Every screen maps to a real field or action
  3. 03

    Prompt, refine, repeat

    No module came out right on the first prompt. Each one took several crafted prompts to get the result we wanted, all built as one interactive prototype on the UpStart design system.

    8 modules · many prompt rounds each · 1 prototype
    1. 01Moderation Queuev1 → vN
    2. 02Q&Av1 → vN
    3. 03Aspect Groupsv1 → vN
    4. 04Question Groupsv1 → vN
    5. 05Configurationv1 → vN
    6. 06Audit Logv1 → vN
    7. 07Dashboardv1 → vN
    8. 08Syndicationv1 → vN
  4. 04

    Build it, then fine-tune it

    Claude Design turns the prompts into working screens with real interactions and realistic data. Then come many small adjustments from a business point of view: wording, rules, states and edge cases that only make sense if you know the product.

    Ratings & Reviews · Claude Design
  5. 05

    Audit every build against the code

    AI fills gaps with confident guesses. After each build I check the output against the code and remove anything it invented, like fields with no backend support or personal details showing on review cards.

    Audit · build vs. code
    • −Fields the backend doesn't support
    • −Personal details on review cards
    • −Anything else Claude Design invented
    • +Only what the API really supports
  6. 06

    Demo, approve, log

    I present the prototype to stakeholders and the client in a demo call. Changes come back as briefs and follow-up presentations. Every decision, prompt and open question is logged in a shared Obsidian vault.

    Demo callStakeholders and client✓ Approved
    Team brain · Obsidian
    • Decisions
    • Prompts
    • Open questions
02 — A real prompt

Channels Management

It started from a product spec and an earlier prototype. I redesigned the consumer model, and researched how Google Merchant, Meta and Algolia work on the receiving end before writing a word of the prompt. Below is one of the many prompts behind it.

Prompt · Claude Design
Update the existing Consumers UI in the Channel Setup design. Keep the current design system, fonts, colors, table styles, nav, and the existing Channels list/detail/create views intact except where noted. This is a redesign of the consumer model, not a restyle.

Consumers are no longer user-created. They are a fixed registry of platform-supported integrations that users CONNECT and CONFIGURE. Remove all "create consumer" affordances. Remove "Archive" everywhere — replace with "Disconnect", which returns a consumer to "Not connected" state in the registry.

Sync schedules and activation/deactivation windows no longer live on the channel. They live on the CONNECTION between a channel and a consumer. Each consumer type defines its own allowed sync cadences:
- Google Merchant: Daily / Weekly / Monthly + "Sync now" (no 15m or hourly — Google fetch schedules don't support it)
- Catalog Service, Web Storefront: Every 15 minutes / Hourly / Daily
The cadence picker must only show that consumer's allowed options.
Researched first
  • Google Merchant
  • Meta
  • Algolia
  1. Redesign the model, not the lookConsumers become a fixed registry of integrations instead of objects users create.
  2. Connect and DisconnectThey replace Create and Archive, so nothing is ever lost.
  3. Rules from the real worldSync schedules live on the channel–consumer connection, and each consumer only offers the cadences its platform supports.
The result · built and iterated in Claude Design
When there’s no spec

Question Groups had no written spec. Only a dormant database schema.

So I designed it purely from the code. Reading the backend first turns a missing spec into a design brief.

03 — This site

This portfolio runs on the same loop

Designed in Claude Design, built and updated through prompts in Claude Code. Every change goes to a Vercel preview first, then live. This is its real build history.

git log · osamajilani.com
  1. Designed in Claude Design and exported
  2. Built with Claude Code, live on osamajilani.comlive
  3. Analytics with Microsoft Claritylive
  4. Lab section with prototype videospreview → live
  5. Site icon, share cards, structured data, sitemaplive
  6. Contact popup, restyled twice on previewpreview → live
  7. PIM case study: screens read straight from Figmapreview → live
  8. New hero line, three versions triedlive
  9. SEO: titles, alt text, IndexNowlive
  10. This pagepreview
  • < 24hfrom design export to live on the domain
  • 22changes shipped since
  • 4features tested on preview links before going live
04 — Prototypes

Built this way

Two enterprise modules for UpStart Commerce, a week each, plus two e-commerce builds.

I use AI tools to turn designs into working prototypes that teams can click through, test and review with real data, long before development starts.

  • Claude Code
  • Claude Design
  • Cursor

Channels Management

1 week

Connect sales channels like Amazon, Walmart and Meta, map product data to each one, and see which listings need attention.

  • Claude Code
  • Cursor
UpStart Commerce · Scala Teams

Ratings & Reviews

1 week

Moderate customer reviews with AI checks, request edits from shoppers, and manage review questions per product.

  • Claude Code
  • Cursor
UpStart Commerce · Scala Teams

LAME storefront

3 hours

A streetwear store landing page with mega menu, product drops and a slide-out cart.

  • Claude Design
  • Claude Code

Watch Winders

2 weeks

A luxury watch store taken from prototype to a live Shopify store: brand pages, collections, product pages and cart.

  • Claude Design
  • Claude Code
Read case study
Say hello

Like what you see? Let’s talk.

Have a project, a role, or just a question? My inbox is open.