The Retro: What ClinicLive Really Cost, and What the AI Got Wrong
Fourteen parts ago, this series made a promise: build a real app with AI and hide nothing — every prompt shown, every mistake kept, every dollar counted. ClinicLive is live behind nginx, its commit messages are the prompt log, and the meter has stopped. Time to count. The bill, honestly These numbers come straight from the build log — cumulative API spend as each part of the app landed: What each…
This series promises to build a real app with AI and disclose every step, from prompts to mistakes, along with the actual costs. The final tally reveals a bill of approximately $6.20 for API usage and about an hour of AI-assisted development time for a Blazor Server app featuring PostgreSQL, SignalR surfaces, authentication, a nine-test suite, CI, and deployment artifacts.
Building the same app from scratch would take two to three days. The billing metric tracks the actual token consumption from the API, priced at the rates for the models used. Despite the low API cost, the real expense lies in the human time spent reviewing and correcting the AI-generated code. The mistake log enumerates the AI's errors, grouped by part of the app development.
Common issues include overlooking business rules, timing differences, authentication mechanisms, and scope limitations. The AI excels at generating a scaffold, translating requirements into database schema and entity mappings, creating a test harness, and providing thorough explanations of errors. However, its limitations become apparent when it fails to grasp the nuances of specific domain rules, time zones, authentication, or scope.
This highlights the essential role of human oversight in AI-assisted development, as the AI lacks the contextual understanding necessary to avoid common pitfalls.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
