A Git Submodule Scaffold That Teaches AI Your Full-Stack Codebase Before It Writes a Single Line
Most AI coding tools treat every repo as a blank slate or rely on generic prompts. This workflow bakes a project's real conventions into structured, reviewable artifacts that survive across sessions and models, cutting down on style drift and low-level mistakes in multi-platform codebases.
A full-stack workspace built around a composite git root repository pulls together mobile, backend, and web submodules. Before any AI coding task runs, a domain-init skill inspects build files, dependencies, and real source code to determine the project type and extract live conventions. The output is a set of per-project assets—profile, context map, verification commands, and a traceable report—that subsequent AI tasks load deterministically.
Domain-init operates on four principles: signal priority to avoid misapplying a frontend spec to a backend, evidence-first rules that must trace back to actual code, domain-filtered context so only relevant governance enters the prompt, and human review gates at every stage. It also ships a quality index that points AI toward specific open-source benchmarks for coding style, testing strategy, and security review, learning how those projects structure rules rather than copying their content.
Once initialized, a task entering from the root reads a workspace map, locates the target sub-project, and follows a fixed loading order: AGENTS.md, the harness core, then the project's profile, context map, and verification spec. The claimed result is a two-day turnaround for A/B feature validation that previously took a week.
Encoding project conventions as structured, reviewable artifacts rather than one-shot prompts creates a persistent governance layer that outlives any single AI session or model swap.
The insistence that rules must cite code evidence—and that humans must judge which patterns are conventions versus historical baggage—is a practical counter to the garbage-in, garbage-out risk of extracting rules from poorly designed legacy codebases.
Pointing AI toward external quality benchmarks to learn rule organization, rather than copying their content, is a subtle but important design choice that avoids polluting a project with foreign conventions.
The claim of compressing a one-week validation cycle into two days is aggressive but plausible if the harness eliminates the manual context-switching and style-correction overhead that dominates multi-platform development.