AI Dev Workflow

Problem: Requirements still drift into code

Many LLM coding flows skip the engineering checkpoints between a rough requirement and production code. Models may miss boundaries, layering, data flow, deployment config, and leave developers to patch the result with follow-up prompts. This skill makes the "think before code" sequence explicit, instead of relying on ad hoc prompting.

Workflow: sample imitation plus prompt-driven rules

It bundles two built-in prompts and reference samples:
- Meta protocol: split complex features into atomic methods; default to method call structure first, with full code on request.
- Architecture prompt: defines layering such as client UI -> ViewModel -> UseCase -> Repository and server Controller -> Service -> Repository -> Entity, with technology variables.
Execution has three steps:
1. Load the feature description sample and expand requirements into requirement -> feature -> execution point.
2. Load the method call structure sample and generate a cross-platform call blueprint for the chosen stack.
3. Generate client, server, database, configuration, and deployment artifacts from that blueprint.

Fit and limits

It works well for small-to-medium features, architecture exercises, and teaching demos where a stable template is useful. For large codebases, strict org standards, or complex domain models, engineers must add constraints manually. The default example stack leans toward Android / SpringBoot / MySQL / Kotlin, so replace it with the project's real path.