My LLM Coding Workflow Going into 2026

Addy Osmani (Chrome Engineering Lead at Google) presents a disciplined, human-directed approach to AI-assisted coding he calls 'AI-assisted engineering.' Rather than treating LLMs as autonomous agents, the framework positions the engineer—or product manager—as the director of a capable but fallible collaborator. The workflow centers on eight interconnected practices: planning first (writing detailed specs before touching code), iterative chunking (breaking work into small discrete tasks to prevent LLM confusion), context provision (supplying relevant code, docs, and constraints proactively), model selection (choosing the right tool for each task type), testing & review (maintaining rigorous quality gates including automated tests and manual review), version control discipline (frequent commits as rollback checkpoints), customization (tuning AI behavior through CLAUDE.md-style rules files), and continuous learning (using AI collaboration to deepen rather than atrophy skills).

The article is particularly valuable for PMs who are themselves vibe-coding or prototyping with AI tools, as it translates engineering-first principles into actionable workflow habits. The 'planning first' approach mirrors best practices for writing AI agent specs—understanding the full problem before delegating to AI prevents compounding errors. The emphasis on human oversight throughout the development lifecycle also gives PMs useful mental models for how to govern AI-assisted work across their teams. Osmani's framing that 'the human engineer remains the director of the show' applies equally to any knowledge worker leveraging LLMs.