Skip to content

AI Vibe Coding Workflow

This page is the authority for the AI-assisted path.

The goal is not to let AI generate only a page. The goal is to let AI generate or refactor a generator that converges toward the platform contract.

The Minimum AI Stack

MCP

Use generator-sdk-mcp so the agent can:

  • read real SDK and runtime documentation
  • inspect available APIs
  • generate starters
  • avoid memory-based guessing

Skills

Use creating-generators for implementation workflow.

When the task is architecture, planning, or refactoring scope, pair it with spec-driven-development and the project-local repository context.

Prompt

A good starting prompt should describe:

  • what the generator creates
  • the preferred stack
  • the required platform capabilities
  • whether template scenarios exist
  • whether the official shell is preferred
  1. configure MCP
  2. load the right skills
  3. let the AI clarify missing inputs
  4. choose starter, shell strategy, and runtime path
  5. generate the runtime base
  6. integrate required SDK modules
  7. add workbench only when the shell should be standardized
  8. verify completion against runtime and platform gates

What AI Should Clarify

Before writing code, AI should ask for:

  • generator identity or appKey
  • stack choice
  • required platform capabilities
  • template-page requirements
  • greenfield vs refactor status
  • compatibility vs standardization target

What AI Should Usually Produce

For a new standard generator, the output should usually include:

  • a runtime starter
  • explicit SDK integration
  • runtime contract surface where needed
  • optional workbench integration
  • honest documentation of unfinished work

Common Failure Modes

  • AI generates only a frontend page
  • AI adds SDK imports but skips runtime standardization
  • AI overbuilds template features without a real scenario
  • AI claims “done” when the result is only a phased refactor

Review Checklist

Before accepting AI output, verify:

  1. MCP was available
  2. the right skills were loaded
  3. AI identified greenfield vs refactor correctly
  4. AI chose the right integration path
  5. runtime contract requirements were handled where needed
  6. remaining gaps were described honestly

Next Step

MIT Licensed