Replatformer

A Warp agent skill that takes websites hosted on managed providers (e.g. Squarespace, Webflow, Wix, WordPress, Shopify) and migrates them into static, source-controlled code that can be deployed to a static-site host such as Vercel, Netlify, or Cloudflare Pages.

Purpose

This skill has two goals:

1. Demonstrate re-platforming — show an end-to-end workflow that moves a hosted site into deployable static code.
2. Demonstrate an outer-loop skill-improvement loop — the skill itself is tuned by an automated outer loop that uses computer use to run the skill against example sites, evaluate the output for correctness and token efficiency, and feed the results back into the skill definition. The point is to show how a skill can be iteratively improved by executing it and measuring the results, rather than hand-tuning it once and leaving it static.

What it does

Given a hosted website, the skill:

1. Inventories the source site — pages, assets, navigation, content, and metadata.
2. Extracts the content and structure into a static site framework (Next.js / Astro by default).
3. Normalizes assets (images, fonts, styles) into the repo so the site no longer depends on the host's CDN.
4. Re-points forms, search, and other dynamic features to serverless equivalents or third-party APIs.
5. Generates a deploy-ready project with the static host's config (e.g. vercel.json), so vercel deploy works out of the box.
6. Verifies the migrated site renders equivalently before cutover.

Why

Hosted providers are convenient but lock you in: limited theming, per-seat pricing, and no access to the underlying code. Re-platforming to a static site gives you full source control, faster builds, cheaper hosting, and the ability to extend the site with code. Beyond the re-platforming workflow itself, this repo is also a demonstration of how to build a self-improving skill: the outer loop uses computer use to run the skill, score the output, and iterate on the skill definition until it is both more correct and more token efficient.

Usage

This repo is an example of a re-platforming skill. Run it from Warp by pointing it at a hosted site URL and a destination framework:

# Example target
./replatform https://example-shop.myshopify.com --to next --deploy-target vercel

The skill walks through the steps above and leaves a deploy-ready project in the current directory.

Project layout

replatformer/
├── README.md                              # This file
├── LICENSE                                # MIT
└── .agents/skills/
    ├── replatform-site/
    │   ├── SKILL.md                       # Inner replatforming skill
    │   └── references/
    │       ├── providers.md               # Provider gathering & feature re-pointing
    │       └── frameworks.md              # Framework scaffolding & deploy config
    ├── replatforming-observer/
    │   ├── SKILL.md                       # Visual/efficiency outer improvement loop
    │   ├── references/                    # Artifact schema and failure taxonomy
    │   ├── scripts/                       # Deterministic metrics aggregation
    │   └── evals/                         # Representative observer evaluations
    └── oz-orchestrated-replatforming/
        ├── SKILL.md                       # Parallel Oz cloud execution driver
        ├── references/                    # Cloud result and handoff schema
        ├── scripts/                       # Deterministic result validation
        └── evals/                         # Representative orchestration evaluations

The skill follows the standard .agents/skills/<skill-name>/SKILL.md layout. The main workflow lives in SKILL.md; provider- and framework-specific detail is split into references/ so the core instructions stay concise.

License

This project is licensed under the MIT License.

Status

The base replatform-site skill, the replatforming-observer outer-loop skill, and the oz-orchestrated-replatforming cloud execution driver are implemented. The observer runs controlled baseline/candidate migrations, compares the source and generated sites with computer vision, measures quality and efficiency, and feeds evidence-backed differential improvements into the base skill. The Oz driver shards observations across parallel cloud agents and gathers cross-site results safely. This repo is published as open source under the warpdotdev-demos GitHub organization.

The original base-skill commit is recorded below so the improvement loops can diff against it as a baseline.

• Base skill commit: f04f3f2 — "Move skill into .agents/skills/replatform-site" (skill in canonical .agents/skills/replatform-site/ location)