What breaks after an AI app prototype
Four small builder projects reached a convincing demo. Their exports then exposed broken run commands, missing runtime configuration, incomplete source handoffs, and demo-only authentication or data. Here is the proof required before launch.
4
Reproduced project cases
6
Production proof layers
$0
Credits, trials, and deployment spend
0
Affiliate links on this page
A successful demo proves the happy path, not the handoff.
The prototype is ready for production review when a clean checkout can build and start, dependencies are reconstructable, authentication is reviewed, state survives restart, data can be restored, and the same commit can operate outside the builder.
None of those proofs is implied by a working preview, visible source files, or a successful production compilation on its own.
These signals are useful, but each stops before production proof
The main screen renders.
Evidence of progress, not independent operation.
The happy-path interaction works once.
Evidence of progress, not independent operation.
The builder preview keeps its managed services attached.
Evidence of progress, not independent operation.
The source tree is visible or downloadable.
Evidence of progress, not independent operation.
Four demos passed before four different handoff failures appeared
Each case used the same small workflow: a marketing page, protected dashboard, three records, one state change, and a non-secret environment label. Observations are product-specific; they are not a universal ranking.
Generated run path
Demo signal: The exported project built successfully, and its generated Cloudflare Worker artifact could run locally.
Observed failure
The generated npm run preview command referenced dist/server/server.js, which did not exist. The app ran only after the Worker artifact was started directly with Wrangler.
Production risk
A valid build artifact can coexist with a broken documented operating path, leaving the next operator unable to reproduce the handoff.
Proof before launch
Test the documented production-start command from a clean checkout and record any builder-specific artifact or adapter it actually requires.
Runtime configuration
Demo signal: The imported repository kept normal Git history and passed npm run build into a conventional Vite dist directory.
Observed failure
The clean clone failed before rendering because the ignored environment file did not provide VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY.
Production risk
Passing a production build does not prove that another environment can reconstruct the runtime dependencies needed to start the app.
Proof before launch
Run the clean clone with a documented environment-variable inventory and verify auth, row-level security, and the server action independently.
Source and database handoff
Demo signal: The hosted preview reproduced sign-in, three records, an environment label, and a reversible status change.
Observed failure
The downloaded project retained the starter commit while generated application files were uncommitted and untracked. Local startup also required DATABASE_URL and SESSION_SECRET, which the export did not include.
Production risk
A working cloud workspace can export source without proving a complete repository round trip, portable database startup, or recoverable production data.
Proof before launch
Commit and push the generated state, then start it locally against an explicitly provisioned database and complete an export-and-restore drill.
Authentication and durability
Demo signal: The generated branch built with the webpack path and reproduced browser sign-in, three records, and a restored API toggle round trip.
Observed failure
Credentials were hard-coded, records and sessions lived only in memory, and a millisecond duration was supplied as cookie seconds, yielding roughly 1,000 days instead of 24 hours.
Production risk
A convincing functional demo can still contain insecure identity logic, non-durable state, and unit errors that only appear during code review or restart testing.
Proof before launch
Replace demo authentication and in-memory stores, verify cookie expiry and revocation, restart the runtime, and confirm state and sessions behave as intended.
Replace “the prototype works” with six separately testable claims
Complete source handoff
Does the repository contain the actual generated application state?
Clone a named commit into a clean folder and confirm generated files, history, branches, and required lockfiles are present.
Documented build and run path
Do the documented commands produce and serve the intended artifact?
Run install, production build, and production start exactly as documented before adding repair steps.
Runtime configuration
Can another environment reconstruct everything the app expects?
Inventory environment-variable names, generated services, callbacks, jobs, and external dependencies without copying secret values.
Authentication and sessions
Are credentials, authorization, expiry units, and session storage production-safe?
Review the generated auth path, test expiry and revocation, and replace demo credentials or in-memory sessions.
Durable data and recovery
Does state survive a restart and can it be restored elsewhere?
Restart the runtime, export records and files, restore them into a clean target, and compare identities and relationships separately.
Independent operation
Can the same commit deploy, roll back, and recover outside the builder?
Deploy to a second target, exercise health and rollback paths, and record ownership of DNS, secrets, storage, and background work.
The cases are narrow by design
The useful conclusion is the test sequence, not a claim that all generated applications fail in the same way.
These four cases do not establish that every AI app builder produces the same defects.
The v0 Turbopack build restriction observed on the test host is not presented as a product-wide build failure.
No non-default public deployment, complete database migration, or authentication-identity migration was completed in this pass.
Webflow remains outside the reproduced case set because its relevant export paths exceeded the zero-spend evidence budget.
Run the six checks before the next production milestone
The free checklist turns these proof layers into controlled choices and returns the smallest evidence gaps to close next. It stores no project name or free text.
Reproduced failures plus first-party operating documents
Observations came from private free-tier projects, repository exports, clean local builds, browser behavior, and reversible state checks. Product documents define the intended Git, runtime, deployment, and data boundaries.
No affiliate link appears on this page. The cases were selected because they were reproduced, not because a program pays a commission.