Mastering AI‑Assisted Development: A 5‑Step Workflow to Keep Full Control

Why AI‑generated code can break systems

Developers often let AI write code directly from a requirement, then spend hours fixing bugs because the AI ignores existing architecture such as caching layers, database migrations, or API contracts. The code may compile and pass tests but becomes a hidden “time bomb” when integrated.

Core principle: Review the plan before any code is generated

Never allow an AI to produce a line of code until the written design has been reviewed and approved. This separates architectural decision‑making from mechanical code generation, ensuring the AI works within explicit boundaries and reduces token usage.

Five‑step pipeline that injects architecture decisions into the AI workflow

Research – Instruct the AI to analyze the relevant parts of the codebase and output a research.md document that captures module dependencies, call chains, and critical constraints.

Plan – Based on the research, have the AI produce a detailed plan.md containing implementation ideas, code snippets, file paths, and trade‑offs. Store the markdown file in version control.

Annotation loop – Add inline comments to the plan to correct assumptions, reject unsuitable solutions, and inject domain knowledge. Send the annotated document back to the AI with a directive to process all comments and update the plan without implementing anything yet.

Task breakdown – Convert the refined plan into a fine‑grained Todo List (e.g., “modify src/api/list.js pagination”, “add type definitions for new endpoint”). The list serves as a progress tracker.

Implement – Issue a standardized implementation command that tells the AI to execute the Todo List, avoid unnecessary comments or any / unknown types, continuously run type checks, and never stop until all tasks are marked complete.

Implementation stage: the AI becomes a mechanical worker, you become the supervisor

During execution, the developer’s role shifts to rapid feedback: point out missing functions, mis‑placed UI elements, or incorrect HTTP methods with concise statements. If the AI drifts off course, roll back all changes and re‑issue a narrowed directive rather than patching incrementally.

Practical tips

Template the research and implementation prompts for reuse across projects.

Persist research.md and plan.md in version control as project assets.

Place annotations directly next to the relevant sections in the plan to give the AI precise context.

Keep feedback during implementation minimal—state only the core issue.

Provide the AI with reference code from the project or open‑source examples to anchor its output.

If a direction proves wrong, roll back immediately and narrow the scope.

Final takeaway

Deeply understand the system, write a concrete plan, iteratively annotate until the plan is solid, then let the AI execute the whole thing while continuously checking types—this separates thinking from typing and keeps developers in control of the architecture.

Original source: https://boristane.com/blog/how-i-use-claude-code/