Boris Tane Reveals Unorthodox Claude Coding Workflow That Prioritizes Planning Over Prompting

Developer Boris Tane has published a detailed account of his unconventional workflow using Anthropic's Claude Code, challenging conventional AI-assisted development practices after nine months of rigorous testing. Unlike typical prompt-and-implement approaches, Tane enforces a strict separation between planning and execution phases, arguing this prevents architectural drift and wasted effort.

"Most developers type a prompt, sometimes use plan mode, fix the errors, repeat. The more terminally online are stitching together complex agent workflows," Tane observes. "The results completely fall apart for anything non-trivial."

Tane's methodology centers on three disciplined phases:

1. Deep-Dive Research
Every task begins with explicit instructions for Claude to thoroughly analyze relevant code sections, requiring findings documented in persistent research.md files. Commands specify intense scrutiny: "Understand deeply... intricacies... don't stop until all bugs are found." This written artifact becomes Tane's verification surface before any planning occurs.

"Without explicit depth signals, Claude skims," Tane explains. "Garbage research means garbage implementation—the costliest failure mode isn't syntax errors, but changes that break system integration."

2. The Annotation Cycle
After Claude generates an implementation plan (plan.md), Tane reviews it locally and adds inline annotations—from brief corrections ("not optional") to architectural redirections ("this should be PATCH, not PUT"). Claude iteratively updates the document through multiple cycles (typically 1-6 rounds) with explicit "don't implement yet" guards preventing premature coding.

"The markdown file is shared mutable state," Tane emphasizes. "I point precisely where something's wrong rather than explaining everything in chat messages. This injects my judgment about product priorities and engineering tradeoffs that Claude lacks."

3. Mechanical Implementation
Only after plan approval does Tane issue a standardized implementation command: "Implement it all... mark tasks completed in the plan... don't stop until done... continuously run typecheck." During execution, Tane shifts to supervisory mode with terse feedback ("wider", "move this to admin app"), using existing code as reference templates and reverting entire changes when necessary.

Key differentiators:

• Persistent artifacts over transient chat history
• Reference implementations leveraged for concrete guidance
• Single extended sessions utilizing Claude's auto-compaction
• Active scope management ("remove this feature from the plan")

"Claude handles mechanical execution while I make judgment calls," Tane states. "The creative work happens in annotation cycles—implementation should be boring."

Developers can explore Tane's complete workflow documentation and example prompts on his technical blog. The approach offers a structured alternative to agent-based systems, prioritizing architectural control through rigorous planning validation before any code generation occurs.