Ship Real MVP

Personalized Codex skills for turning vague ideas and visual references into working personal software.

Comment

Inspiration

Codex can already turn an idea into working software remarkably fast. The harder problem is staying faithful to what the user actually meant: avoiding invented requirements, calibrating visual references correctly, verifying the real product instead of only the build, and knowing when human judgment is still required.

I built Ship Real MVP to explore whether personalized Codex Skills could make that process more controlled and more verifiable without growing into another heavy framework.

What it is

Ship Real MVP is a pair of reusable, personalized Codex Skills:

  • Ship Real MVP 02 guides risk-driven discovery, narrow vertical-slice delivery, recovery states, verification, and a five-lens PM / UI / UX / Development / Testing review.
  • Adapt UI From References Mini 02 separates multiple visual references by layer ownership, rejects source-product semantics that should not transfer, calibrates one high-fidelity core state before expansion, and applies a Realness Gate against colored-wireframe results.

The reusable project is the development method. A working SwiftUI camera app, lilt Style Camera, demonstrates the method in practice.

How the benchmark worked

I used the same short Chinese product request, the same Instagram color reference, the same Pinterest UI reference, and Codex with GPT-5.6 Sol at medium reasoning effort in two branches:

  • Baseline: general Codex capabilities and general-purpose Skills, without the two personalized GapLab Skills.
  • Skill-guided: the same setup with both personalized Skills enabled.

I responded as the intended user: I expressed real dissatisfaction and preferences, but did not provide technical implementation solutions. The two conversations were allowed to evolve naturally rather than being forced into identical transcripts.

What happened

The Baseline reached a broad MVP quickly. When the visual reference was initially inaccessible, it invented three presets that were never requested. Its first physical-device capture crashed. After correction, MUSE 01 produced creamier highlights that I preferred in some respects, but UI spacing and interaction stability remained weaker, and the latest capture patch was not fully retested on device.

The Skill-guided path explicitly assigned:

  • Instagram to color and texture
  • Pinterest to UI material and layout
  • camera behavior, permissions, saving, and iOS semantics to the target platform

Codex paused after one high-fidelity core state for calibration. When I said the color still felt wrong, it analyzed more frames, continued implementation, and performed a five-lens review.

The resulting lilt app completed the real capture-and-save flow on a physical iPhone. During that session I observed no crash, blocking bug, severe lag, or unresponsive control. The project still records the unresolved truth: I did not fully accept the final color style or the post-capture confirmation flow, and no Release Archive or App Store submission was completed.

How Codex and GPT-5.6 were used

Codex with GPT-5.6 Sol inspected and routed references, created SwiftUI projects, implemented AVFoundation and Core Image pipelines, generated tests, captured simulator evidence, analyzed failures, produced release documentation, and reviewed the product from multiple perspectives.

The personalized Skills changed when Codex stopped, what it asked the user to evaluate, how evidence was classified, and how completion claims were constrained. The human remained responsible for taste, physical-device experience, and publishing boundaries.

What I learned

There was no universal winner.

Default Codex was faster and sometimes more creatively surprising. The personalized Skills improved control, UI-reference fidelity, evidence discipline, and review coverage, but added interaction cost and could refine a wrong early direction too deeply. More analysis did not guarantee better subjective taste.

The strongest conclusion was:

Default Codex builds fast. Personalized Skills make it stop, look, and check—but they must not become another framework.

How to test

Judges can test the two Skills without rebuilding the iOS app. The public repository contains installation instructions and a deterministic read-only audit prompt in docs/submission/JUDGE_TEST_GUIDE.md.

The included iOS example can also be opened in Xcode. Simulator use is suitable for UI inspection; a physical iPhone is required for real camera capture.

Repository and evidence

The repository includes both Skills, the working lilt example, the Baseline contrast app, the shared protocol, comparison report, verification evidence, setup instructions, supported-platform notes, and known limitations.

The app is the output. The reusable project is the way it was built.

Built With

Share this project:

Updates