ClaimDone

Inspiration

After a stressful incident, people must collect evidence, remember important details, and translate an ordinary explanation into an unfamiliar insurance form. ClaimDone makes that first step feel as simple as sharing a few photos and explaining what happened.

What it does

ClaimDone turns one to three accident photos plus a short text or voice description into a clear insurance-claim. The multimodal claim agent reviews the visible evidence and the customer statement. Once the claim is complete, Computer Use opens a dummy insurer website, navigates from its home page to the claim form and fills only the approved claim values.

How I built it with Codex and GPT-5.6

The runtime AI pipeline uses:

• GPT-5.6 for multimodal image and statement analysis through the OpenAI Responses API
• gpt-4o-mini-transcribe for voice memos
• gpt-5.4-mini Computer Use for the restricted insurer-portal handoff

Codex was my used for all of the development. I used it to plan the product flow, implement the frontend and backend, create the restricted Computer Use workflow, review the responsive interface in a browser, diagnose failures, write tests, and prepare the repository and demo. GPT-5.6 Sol extra high was used for deep planning. GPT-5.6 ULTRA for milestone-based goals, and GPT-5.6 Terra supported fast iteration and day-to-day questions.

A key workflow lesson was to define the product in Plan mode first, when I don't know what exactly to build. Then give longer Codex goals with explicit milestone stops. After every milestone I reviewed the working experience, gave feedback, and continued after approval. That produced a much smaller and more coherent result than giving a long-running agent an underspecified destination.

Accomplishments I'm proud of

• Real Computer Use navigation and form filling inside a safe synthetic portal
• One simple customer flow with only four states: input, analyzing, needs information, and ready
• Observable agent activity that remains useful without exposing hidden reasoning

P.S.

I learned that Codex works best for ambitious builds when product uncertainty is resolved before implementation. Planning, precise goals, and review checkpoints made the agent dramatically more effective.

Built With

• codex
• computer-use
• gpt-5.6
• next.js
• openai
• openai-responses-api
• playwright
• react
• typescript