How I'd build Goodword
I'd reach for Next.js on the frontend with a FastAPI backend handling async review monitoring, Postgres for multi-tenant data, Stripe for billing, and Claude's API for response generation. The stack is solid for SaaS at this scale. Rough estimate: 320-360 hours, which lands us at production in 8-10 weeks with a full-time focus, or 4-5 months part-time. That covers MVP through initial customer rollout.
Day-by-day plan
- Day 1-2: Set up auth (I'd use Clerk or Auth0), multi-tenant account schema in Postgres, and basic RBAC so business owners can invite staff members.
- Day 3: Integrate Stripe for billing. Three tiers: free trial, $49/month core, $99/month with Yelp/Facebook. Webhook handlers for churn alerts.
- Day 4-5: Build customer onboarding flow. Business name, Google My Business account link, review source selection, response tone/brand voice config.
- Day 6-7: GMB API integration. Set up OAuth flow so we can authenticate to Google, fetch reviews, and handle rate limits (Google caps requests hard). Cache reviews in Postgres.
- Day 8-9: Response generation pipeline. Hook Claude API to draft responses based on stored review data and customer's brand voice settings. Approval staging UI where the owner reviews the draft before we post.
- Day 10-11: Email notifications (Resend) when new reviews drop. Dashboard showing review count, response status, and basic metrics.
- Day 12-14: Yelp assessment and decision gate. Build scraper proof-of-concept to evaluate feasibility, or scope it out of MVP and keep it as a known limitation in the trial.
- Week 3-4: Customer testing, bug fixes, and runbook documentation. Soft launch to 5-10 customers.
What's hard about this build
The core risk is API dependency. Google My Business explicitly forbids automated review responses in their terms - they can revoke our API access in a day, which instantly breaks the product for every customer. I'd mitigate by building a human-in-the-loop approval flow and keeping customer communications transparent about the ToS risk. Yelp is harder: they have no legitimate API for review responses. We can either scrape (violates their ToS and is fragile), or we accept it's not buildable and remove it from the MVP. Facebook's moderation APIs are strict too. The second risk is churn. At 5-8% monthly churn with 200 customers, you're losing 10-16 subscribers every month just from natural attrition. Unit economics don't work unless acquisition is near-zero or you're expanding features like response analytics. I'd track response-to-conversion rates for reviews - that becomes the retention hook.
What's fast because of AI
Claude accelerates the core value prop immediately. Response drafting that a human copywriter takes 15 minutes per review becomes 2 seconds via the API. I'd lean on Claude for prompt engineering edge cases: handling angry reviews, industry-specific language (dental vs. restaurant vs. law firm), multi-language responses. For testing, Claude writes exhaustive test cases for the review-matching logic - detecting duplicates, handling deleted reviews, edge cases in tone detection. Scaffolding API routes, database migrations, and boilerplate billing flows compresses weeks into days. The initial runbook and troubleshooting guide comes faster too. The slowest parts are still the regulatory and ToS diligence and customer research, which AI can't shortcut.
How I'd hand it off
I'd leave a Loom walkthrough of the admin dashboard, customer flows, and the approval/posting pipeline. A runbook covering common failure modes: API rate limits, failed Stripe charges, Gmail OAuth expiry, and the escalation path for Google API access issues. A 30-day pager rotation with weekly check-ins on churn metrics and new customer feedback. All credentials (Stripe keys, Google service account, Claude API keys) stored in a shared vault. A Linear board with known bugs, feature requests from early customers, and the Yelp decision (build, scope, or sunset). The codebase is well-commented around high-risk areas: API authentication, multi-tenant isolation, and response approval logic.