Why Massive Prompt Dumps Break Codebases: The Case for Branch-Isolated Execution

The most common failure mode in modern AI coding assistants is what we call the "Massive Prompt Dump". You describe a complex feature—say, user authentication with JWT, role-based access, and session management—and the AI attempts to rewrite 8 files in a single monolithic output. Inevitably, subtle syntax errors slip in, imported dependencies are omitted, and existing working functions get silently overwritten. The codebase is broken, and tracing the regression is an exercise in frustration. We realized that human engineering teams don't build software this way. We use **Atomic Tasks and Isolated Git Branches**. In Ashwatth, our AI agent acts as a Staff Engineer and Sprint Lead: 1. **Automated Spec Decomposition**: When a feature spec is approved, the agent decomposes it into granular, layered engineering tasks: - `[design]`: Canvas-only mockups and UI tokens. - `[build] - Data Layer`: Postgres SQL migrations (`/schema.sql`) and TypeScript/Zod schemas. - `[build] - API Layer`: Express and Next.js App Router route handlers with strict request/response contracts. - `[build] - Frontend Wiring`: Connecting approved UI components to real backend endpoints. - `[fixes]`: Loading skeletons, error toasts, and edge case verification. 2. **Branch-Isolated Execution**: Each Kanban task is executed on a dedicated Git branch with localized AST updates (`replace_code_lines`, `write_new_file`). 3. **Automated QA Verification**: The host dev server compiles the branch. Only after linting and runtime execution checks pass is the task marked complete. By enforcing atomic, layered task execution, regressions drop to near zero and developer trust is restored.