What is replica-skill? Building a reading-list app prototype
An app screenshot does not reveal all the error states, persistence behavior, or user flows you need to build. Replica organizes these into a sequence. We applied selected steps to a small reading-list app to examine how handoff artifacts and test criteria expose unfinished work.

Replica divides app-building work across 11 Skills
Jakeschincariol/replica-skill provides Agent instructions for reconnaissance, architecture, UI, backend work, testing, comparison, branding, and deployment. It is not an AI model or a one-click app ready for sale.
The source calls for studying functionality and user flows, then writing a fresh implementation without taking the original app's code, logo, images, or copy. Data rights and source terms still need to be assessed for each project.
What each stage passes to the next
replica-recon produces a screen map, user flows, and a feature matrix. Next, replica-architect plans the architecture, and replica-design defines colors, spacing, typography, and components before replica-build implements the plan.
Meanwhile, replica-test and replica-diff check behavior and identify missing features. Backend, entrepreneur, brand, launch, and deploy stages handle subsequent work as scoped. The pack's value lies in inspectable handoffs, such as a feature matrix that can mark work partial instead of calling everything done.
Testing the reading-list app's core loop
KrobClick used the task-list concept from TodoMVC as a public reference brief, then wrote a new Thai reading-list app. We followed the recon, architect, design, build, and test instructions in Codex without installing the Claude plugin or running Replica's Python scripts.
The scope includes adding titles, changing read status, filtering, deletion, and localStorage persistence. It excludes accounts, synchronization, title editing, and a backend, and does not aim for complete TodoMVC feature parity. The plans, feature matrix, code, and checks are stored with this article's work for reproduction.
What the prototype checks establish
The checks passed 11 groups of cases covering empty lists, whitespace-only input, HTML-looking text, long Thai titles, filtering, deletion, persistence after reload, keyboard use, corrupt saved data, denied writes, and the absence of page errors or external requests. At widths of 320, 393, and 1440 pixels there was no page overflow, and axe found no violations in the checked states.
These results establish the behavior of this prototype in the cases actually run. They do not certify accessibility in every condition, multi-tab behavior, or service readiness. We did not compare completion time against work without the Skill, so we cannot quantify time savings.

Check that data is saved and failures are reported
One consequential case occurs when the browser refuses localStorage writes. The list can still work on the current page, but must warn that it may disappear after reload. A polished screen that silently looks saved would misrepresent its behavior.
Input containing HTML syntax must also display as text. We added such a title and checked that it did not execute code. These criteria make “working” more specific than simply loading a page.
Choose one app flow to test end to end
Before using the approach on a real app, choose one small goal, such as adding an item and finding it after a refresh. State what is outside scope and plan checks for empty states, errors, and mobile screens from the start. Add accounts or payments when the task needs them so they do not obscure the behavior being tested.
Once the small loop works, assess the database, permissions, and production requirements. Read How to verify an Agent's work to distinguish a success report from evidence of what actually happened.
Sources
- Jakeschincariol/replica-skill: Documentation for the referenced version
- TodoMVC: Task-list reference brief
- Jakeschincariol/replica-skill
Reviewed Replica commit 77c9436fb3d18c3d58169efb8caf4fe906b0dc51 on October 30, 2026. We followed selected instructions in Codex and tested a real prototype. We did not test all 11 Skills or install the Claude plugin.