When 90% of Code Is AI‑Generated: Why Mastering the Right Workflow Matters

In 2025 AI coding assistants have reshaped software development, but many developers still struggle to turn experimentation into real productivity gains. Addy Osmani, a Google engineering lead, shares a systematic workflow that makes AI a reliable partner rather than a source of chaos.

1. Plan before you code

Instead of giving vague prompts like “write a login feature,” Osmani first collaborates with the AI to brainstorm and produce a detailed specification document (spec.md) covering requirements, architecture decisions, data models, and test strategies. The spec then drives a project plan that breaks the work into small tasks and milestones, ensuring both human and AI stay aligned and reducing rework.

2. Break work into tiny, iterative steps

Osmani stresses never letting the AI output large code blocks at once. He decomposes the project into bite‑size pieces, letting the AI focus on a single function, bug fix, or feature per iteration. After each step, the result is tested before moving on, preventing the AI from drifting and avoiding “messy code.”

3. Feed sufficient context

The AI’s performance hinges on the context it receives. Osmani performs a “brain dump” before coding, providing high‑level goals, good examples, and pitfalls to avoid. For complex APIs he manually copies relevant documentation or README, and uses tools like gitingest or repo2txt to automatically dump repository contents into text for the AI.

4. Choose the right model and switch when needed

Different tasks benefit from different models. Osmani selects Claude (Anthropic) for strong reasoning and architecture design, Gemini (Google) for natural interaction, GPT (OpenAI) for balanced capability, and tools like Cursor or Copilot for real‑time coding assistance. If a model stalls, he swaps to another.

5. Never trust AI output blindly

He treats the AI as an over‑confident junior developer that must be verified. Every piece of generated code is run through unit tests or manual checks. Osmani often generates a test plan first, or lets a second AI session review the code, adding an extra layer of scrutiny.

6. Commit frequently with fine‑grained version control

Osmani commits after each tiny task, creating “save points” that allow easy rollback if the AI misbehaves. Frequent commits also serve as logs; the AI can parse git diffs to understand context, and reviewers (human or AI) can more easily assess changes.

7. Customize AI behavior with rules and examples

He supplies the AI with a CLAUDE.md or rules file defining code style, linting standards, and other conventions. Custom instructions or inline examples at the start of a session guide the AI’s tone and output, turning it into a disciplined team member rather than a black box.

8. Embrace testing and automation

A robust CI/CD pipeline lets the AI trigger automated tests and receive immediate feedback. Failed tests are fed back to the AI as error messages to fix, creating a tight bug‑fix loop. Linting and type checking act as strict teachers that catch errors early.

9. Continuous learning and adaptation

AI amplifies the developer’s existing skills: solid fundamentals let the AI handle design and architecture, while weak foundations can lead to amplified chaos. The workflow ultimately promotes “AI‑enhanced software engineering” rather than full automation.

Reference: Addy Osmani, “My LLM coding workflow going into 2026”, Substack, 2025

https://addyo.substack.com/p/my-llm-coding-workflow-going-into