Four SaaS prototypes, built by directing AI coding agents

What they have in common

AI coding agents can produce a working-looking app in an afternoon. What usually decides whether software survives production is the part they tend to skip: what happens when a key is missing, a provider is down, or one customer’s data is a single bad query away from another’s.

All four apps are built to run with as little configuration as possible and degrade honestly when an integration is missing — never crash, never pretend. LeadParrot and ServiceQuote AI start with an empty environment file; the other two need only a Supabase project.

What they are

• LeadParrot finds public posts from people asking for what a business sells, scores buyer intent and drafts a reply for a human to review. It never posts or messages anyone automatically.
• ServiceLead AI texts back a caller whose call was missed, runs a short AI intake and hands the owner a scored lead summary.
• ReviewLoop AI turns completed jobs into review requests and drafts replies to reviews. The owner approves everything.
• ServiceQuote AI turns a contractor’s job notes into an editable, line-item estimate against the contractor’s own price book.

Patterns they share

• Provider abstraction for AI: a deterministic mock by default, with OpenAI or Anthropic when a key is present.
• Organisation-scoped data with Postgres row-level security, so one customer’s data cannot leak into another’s through a missed filter.
• Every optional integration degrades to a mock or a setup notice: payments show “not configured”, email falls back to an in-app preview, webhooks refuse unverified events once a secret exists.
• Humans stay in control: nothing is posted, sent or priced without the owner approving it.

What these do and do not show

They show how I specify and review agent-built software, and the defaults I insist on. They do not show production operation: none was deployed, none has users, and each was built in a few days. My production work is in the other case studies.