How I'd build Relay
I'd go with Next.js on the frontend with Postgres on the backend, Stripe for billing, and the Figma and GitHub APIs for artifact ingestion. For document generation, I'd use Puppeteer for PDFs and the Notion API for Notion exports. I'm estimating 140 hours total, so roughly 1.5-2 weeks of solid build time. That's realistically three weeks with testing and deployment.
Day-by-day plan
Day 1: Set up Next.js project, Postgres schema for projects, users, and generated documents. Multi-tenant isolation by user_id. Day 2: Stripe integration with three tiers (free trial, freelancer, agency). Subscription status checks on protected routes. Day 3: Figma OAuth, artifact inventory logic. I'd parse layers and component libraries, flag unpublished components and placeholder frames. Day 4: GitHub API integration. Read repository structure, extract README content, identify environment files. Day 5: Prompt engineering for the four core sections. Asset inventory with verified paths, usage guide translated to client language, maintenance playbook with time estimates. Day 6: Support terms generation based on freelancer's hourly rate and declared scope limits. Day 7: Build the in-line editing UI for each section. Users can override generated content before approval. Day 8-9: Export pipeline. PDF rendering via Puppeteer, Notion page creation, Google Docs optional. Day 10: Full testing, edge case handling, security audit before launch.
What's hard about this build
The biggest technical risk is LLM hallucination in client-facing documents. If the agent invents a file path or fabricates a support term, that goes straight to the client and kills trust immediately. I'd mitigate this by having the agent only reference artifacts it actually verified, flag any inferred content, and require freelancer sign-off before export. The second risk is Figma layer parsing. Component libraries with inconsistent naming, deeply nested frames, and unpublished variants can confuse the inventory logic. I'd need to validate against real Figma projects early. Third, multi-source-of-truth handling gets messy fast. A user with both a Figma file and a GitHub repo might have conflicting information. I'd surface conflicts to the user and let them disambiguate rather than guessing.
What's fast because of AI
Scaffolding the Next.js API routes and database migrations compresses what's normally two days into one afternoon. Claude can generate the Prisma schemas, the endpoint structure, and the Stripe webhook handlers in one session. Writing the prompts for section generation is where Claude really earns its keep. I'd iterate on drafts with real Figma files and GitHub repos, and Claude's feedback on prompt clarity saves a full day of testing. Test generation is another huge win. Instead of writing 50 edge cases by hand, I feed Claude the function and ask it to enumerate failures, then generates the test suite. Debugging integration failures with the Figma and GitHub APIs is also accelerated. Claude can read the API docs, spot what I'm missing, and suggest fixes faster than I could hunt through documentation.
How I'd hand it off
I'd record a Loom walkthrough of the admin console, the payment tier logic, and how to configure Figma and GitHub OAuth. I'd write a runbook for common incidents: what to do if Stripe webhooks stop, how to rollback a bad deployment, how to manually trigger a document export if the queue stalls. I'd stay on pager rotation for the first 30 days at no extra charge, responding to alerts within an hour during business hours. I'd transfer Stripe credentials, API keys, and Figma OAuth app secrets to your 1Password vault. Database backups run nightly to S3. I'd leave the codebase in Linear with all open issues documented and next priorities flagged.