Terminal
Terminal/Skill
Translate a source application into a faithful web prototype through discovery, design and logic capture, gap planning, iterative implementation, and evidence-based review.
By Builder.io
Works with
npx skills add builderio/builder-agent-skills --skill web-prototypeReplicate a non-web source application in an existing web prototype. Keep durable process files in .agents/skills/web-prototype so later runs can resume the work and improve the translation.
Complete this phase before changing application code.
Look for preferences.md in .agents/skills/web-prototype. If it exists, read it and verify that the recorded paths still exist and match their recorded project types. Use valid saved preferences instead of repeating discovery.
Search the repository for non-web application projects whose interface and behavior could be replicated. Candidates include:
The source is the specific project or subdirectory that acts as the product source of truth. Do not select generated files, build output, dependencies, or the entire repository when a narrower application directory is available.
Search the repository for projects built with a web framework. Detect them from manifests and framework configuration such as package.json, angular.json, Next.js, Nuxt, Vite, SvelteKit, or similar files.
The prototype is the specific web project or subdirectory where the source experience will be replicated. Exclude static documentation, dependency directories, generated output, and unrelated websites.
Verify that the source uses a non-web framework and the prototype uses a web framework. If there is exactly one credible source and one credible prototype, select them. If multiple candidates exist, briefly explain each candidate and ask the user which pairing to use. Never guess between multiple applications.
Create or update preferences.md in .agents/skills/web-prototype with confirmed repository-relative paths:
# Web Prototype Preferences
- Source: `<repository-relative path>`
- Source framework: `<native framework or platform>`
- Prototype: `<repository-relative path>`
- Prototype framework: `<web framework>`
Replace stale values when the user confirms a different pairing.
Inspect the complete source and compare its available experiences with the prototype. Do not infer coverage from filenames alone: read implementations, trace navigation and state, and run the source when the environment permits.
Create or update design.md in .agents/skills/web-prototype as the canonical inventory of everything visual in the source. Preserve concrete file or symbol references so each finding is traceable. Capture at least:
Describe observable behavior and relationships rather than copying framework-specific implementation details. Clearly mark anything that could not be inspected or verified.
Create or update logic.md in .agents/skills/web-prototype as the canonical inventory of source behavior. Preserve concrete file or symbol references and capture at least:
Separate business requirements from native-framework plumbing. Record uncertain or unreachable behavior explicitly instead of inventing it.
Compare every entry in design.md and logic.md with the current prototype. Create or update work.md in .agents/skills/web-prototype as a thorough list of only the outstanding translation work.
Organize the list by page or user journey, then by design and logic. Every work item must include:
design.md or logic.md section and source referenceCover all components, pages, routes, states, responsive behavior, tokens, assets, animations, transitions, and business rules. Split broad items until progress can be reported meaningfully. Do not include completed work, speculative enhancements, native-only platform chrome, or unrelated refactoring.
Before implementation, confirm that every captured design and logic requirement is either already satisfied or represented by an item in work.md.
Read learning.md before planning or changing code. Apply relevant prior learnings throughout implementation.
Work through work.md in dependency order, one coherent item at a time:
work.md only after implementation and verification are complete.If implementation reveals missing or incorrect source knowledge, update design.md or logic.md, add the resulting gaps to work.md, and continue from the corrected documents.
Create or update learning.md in .agents/skills/web-prototype whenever the work reveals a reusable lesson. Record durable facts such as effective translation patterns, repository conventions, source-to-web mapping decisions, testing techniques, mock strategies, and pitfalls. Do not use it as a chronological activity log. Never store secrets, credentials, or personal data.
Make prototype changes only inside the recorded prototype project unless the user explicitly requests otherwise. Do not modify the source while translating it.
Perform a final source-to-prototype review after the work queue appears complete. Use design.md, logic.md, learning.md, and the source itself as evidence rather than relying on memory.
Walk every captured route and state in the browser. Confirm that the prototype accurately represents the source's components, pages, content hierarchy, layouts, responsive behavior, design tokens, assets, visual states, animations, gestures, and transitions. Compare at representative viewport sizes and exercise navigation rather than reviewing isolated screenshots only.
Demonstrate each captured journey and business rule through tests, deterministic fixtures, or visible prototype controls and states. Confirm calculations, validation, state transitions, persistence behavior, errors, retries, and conditional flows. Verify that external data and authentication are intentionally mocked while preserving the expected contracts and user-visible outcomes.
Reconcile every section of design.md and logic.md against the prototype. Add any discovered gap back to work.md and return to Phase 4. Remove verified items from work.md; do not retain completed checklist entries.
The process is complete only when work.md has no outstanding items, every design and logic requirement has review evidence, the prototype works in the browser, and relevant automated tests pass. Report any environment limitation that prevented direct verification instead of claiming completion.