← Ecosystem

Loptr Lab · Preliminary case study · October 26, 2026

The hidden coordination costs of AI-assisted development

AI made it possible to produce prototypes, tests, documentation, accessibility repairs and creator resources without immediately funding a full development team. But it also enabled overlapping sites, repeated correction cycles and unclear ownership of source and deployments.

What was delivered

Across separate reported checkpoints, Made-Sick passed 76 automated tests, 35 local browser checks and 18 live read-only checks. PIXIE reconciliation reported 43 passing tests. These numbers are not additive and do not establish full production acceptance. Duet offline recovery, creator tools and public guidance were also developed, while their individual release gates remain distinct.

Where the process failed

Competing entry points, duplicated training publication, inconsistent UX and source-control ambiguities created coordination debt. A ChatGPT Sites implementation proved difficult to retrieve through Codex's internal Git endpoint. The failure was not solely generated code: external services and authentication restrictions also contributed. Repeatedly troubleshooting a known blocker without an effective stop rule was avoidable.

Costs versus a human-led team

An LLM subscription is less expensive than a full development salary, but they do not purchase equivalent services. The correct denominator is accepted work per total human attention, including prompting, inspection, repairs, integration, QA, deployment and oversight. Exact LLM spending, founder hours and counterfactual human-team performance were not audited; salary and rework examples in the longer editorial draft are illustrative, not measured savings.

Corrective standards

1. Check existing work before introducing a new site, repository, guide or dependency.
2. Obtain explicit owner approval for platform, scope and release changes.
3. Keep GitHub source, review branches and deployment authority documented.
4. Preserve distinct project brands, but make paths through the ecosystem coherent.
5. Test accessible, real end-to-end journeys before claiming completion.
6. Stop repeating failed authentication attempts until access changes.

What remains unverified

Complete source synchronization for certain Sites histories, some authenticated account journeys, independent assistive-technology and physical-device acceptance, and actual all-in AI development costs. This report does not represent those as complete.