How I'd build Standing
I'd build this as a Next.js full-stack app on Vercel with a Postgres backend, Stripe for billing, Resend for email, and Plaid to ingest freelancer revenue history if they want automated client scoring. The core agent runs on Claude's API with prompt caching to keep proposal drafting cheap at scale. I'm estimating 110-130 hours to production: data modeling, auth, the proposal engine, email tracking, and basic reporting. This assumes you're not building legal document generation yourself - I'd use a headless DocuSign or PandaDoc integration for the agreement step.
Day-by-day plan
- Day 1-2: Multi-tenant auth with Clerk, Postgres schema for freelancers, clients, projects, billing records, and proposal versions. Wire up Stripe Customer creation at signup.
- Day 3: Build the client intake form - file upload (CSV or Stripe API sync), data normalization, and initial deduplication.
- Day 4: Implement the proposal scoring engine in Python (FastAPI sidecar or Vercel Functions) that consumes the client list and calculates retainer readiness using heuristics: frequency of work, project recency, rolling dollar volume.
- Day 5: Draft the proposal generation endpoint. Feed client context plus freelancer voice samples plus recommended scope into Claude with prompt caching. Render as a template (HTML email plus web preview).
- Day 6: Email delivery (Resend) with click and open tracking via Postmark webhooks. Store events in Postgres. Build basic dashboard showing proposal status per client.
- Day 7: Implement follow-up nudge sequence (scheduled with Bull or pg-boss) and client acceptance flow - simple "accept retainer" button that sets up billing.
- Day 8-9: Stripe recurring billing wiring. Auto-generate first invoice, attach agreement preview, hand off to DocuSign API for wet signature if requested.
- Day 10: Admin reporting, basic churn alerts, and a 30-day onboarding checklist for the freelancer.
What's hard about this build
The proposal engine has to capture the freelancer's voice accurately without hallucinating - we'll need to feed Claude multiple past email samples and score generated proposals against a rubric before sending. The real gotcha is client deduplication across Stripe, CSV uploads, and Plaid data; you'll have false positives on "John Smith" and miss "John S. Smith Inc." I'd use a fuzzy match library and hand-verify the top 20 clients before scoring runs. Email deliverability is also load-bearing: if proposals land in spam, activation drops. I'd implement DKIM and SPF setup checks at onboarding and use Postmark's template system to keep sending clean. The legal step (agreement generation) isn't just technical - DocuSign has a complex signing flow and you need clear copy around what happens if a client refuses to sign.
What's fast because of AI
Claude handles 70 percent of the proposal generation workload. I'd use prompt caching on the freelancer's past communication samples and a cached context block on Standing's tone guide, so subsequent proposals for the same freelancer are nearly free. Test generation is a huge accelerator: I'd ask Claude to enumerate edge cases (client with zero recurring projects, freelancer with 200 past clients, proposal sent to a nonresponsive contact), then scaffold the test suite from those cases. Copy for the onboarding flow and dashboard takes a day normally; Claude writes it in 2 hours, I refine it in 2 more. Debugging is where Claude excels - instead of staring at logs, I describe the behavior, Claude traces through the logic, and we're unstuck in 15 minutes.
How I'd hand it off
I'd record a Loom showing the full freelancer journey: uploading a client list, reviewing proposal drafts, sending, and activating the first retainer. You get a runbook covering deployment steps, Stripe webhook configuration, Resend API keys, and how to handle a DocuSign integration flip-flop if the freelancer opts out of agreements. I'd join a 30-day pager rotation to field production issues - payment failures, email bounces, Plaid sync errors. All credentials move into your vault. The codebase goes into Linear with 12 months of commit history so the next operator has context.