Developer Boris TaneBoris Tane's approach to using Claude Code has been making waves in the developer community. The workflow he refined over nine months of experimentation offers a completely different approach from the conventional one of simply asking an AI to write code. It delivers an important insight into the role we should play in the human-AI collaboration of the future.
The Limits of Conventional AI Coding—and a New Approach
The problem most developers run into when they use AI coding tools is clear. Writing prompts on the fly, fixing the errors, and repeating the cycle works well for simple tasks but falls apart completely on complex projects. The core of Tane's solution is simple: never let it write code until you have reviewed and approved the plan.
What makes this approach revolutionary is that it fully separates planning from execution. Humans take charge of strategic thinking and judgment, while the AI handles mechanical execution—a clear division of labor. This minimizes token usage while producing far better results.
A Disciplined Three-Stage Workflow
Stage 1: Thorough Investigation and Research
Every task begins with a 'deep-readdeep-read' instruction. You ask Claude to thoroughly understand the relevant parts of the codebase, but you require it to write up its findings in a Markdown file. You have to explicitly use words like 'depth,' 'detail,' and 'complexity' so the AI doesn't settle for a surface-level understanding.
This research stage matters because it heads off the most expensive failure mode in AI coding. More fatal than a syntax error or a logical flaw is an implementation that works in isolation but breaks the system as a whole. It pre-empts problems like a function that ignores an existing caching layer, or a migration that fails to account for the ORM's rules.
Stage 2: Drafting the Plan and the Annotation Cycle
After reviewing the research, you ask for a detailed implementation plan in a separate Markdown file. The distinctive method Tane uses here is the 'annotation cycleannotation cycle.' Once the AI writes the plan, he adds inline comments directly in his editor.
The comments vary. Sometimes he corrects a parameter's necessity with just two words—'not optional'—and sometimes he writes a long paragraph explaining a business constraint. The key is that instead of explaining it in a chat message, he marks the fix at the exact spot where the problem is.
As this cycle repeats one to six times, a generic implementation plan transforms into one that fits the existing system perfectly. The AI excels at understanding code and proposing solutions, but it doesn't know product priorities, user pain points, or engineering trade-offs.
Stage 3: Mechanical Execution
Once the plan is complete, you issue the implementation command. The standard prompt Tane uses in almost every session goes like this: "Implement everything. As you finish a task or a step, mark it complete in the plan document. Don't stop until everything is done."
At this point, because every decision has already been made, the implementation is mechanical rather than creative. That's by design. The creative work was already finished in the annotation cycle; all that remains now is tedious execution.
A New Paradigm for Human-AI Collaboration
The most important insight this workflow offers is the division of roles between humans and AI in the knowledge work of the future. In an era where AI becomes not just a tool but a collaborative partner, what role should we play?
Tane's methodology gives a clear answer. Humans judge; AI executes. The unique value humans bring is domain knowledge, business context, an understanding of users, and—above all—judgment. Because the AI doesn't know this context, it can propose solutions that are technically flawless yet wrong for the project.
Take solutions that are over-engineered, that change a public API signature other parts of the system depend on, or that pick a complex option when a simple one exists. Humans can make these calls because they hold the context for the wider system, the product direction, and the engineering culture.
Splitting Knowledge into Skills, and the Road Ahead
This way of collaborating becomes the basis for a fresh division of our knowledge and skills. Where a single developer used to handle everything from grasping the concept to implementing it, the roles now break down into specialized parts.
The human's exclusive domain: The 'strategic thinking'—understanding system architecture, interpreting business requirements, judging technical trade-offs, setting code-quality standards, and weighing user experience—belongs to humans.
The AI's optimal domain: The 'executional work'—repetitive coding tasks, syntax checking, pattern recognition, analyzing large codebases, and mechanical refactoring—falls to the AI.
This split will apply not only to development but to other areas of knowledge work as well. In writing, humans will decide the message and tone while the AI polishes the grammar and structure; in research, humans will set the hypotheses while the AI gathers and organizes the data.
Key Principles for Practical Application
The practical principles you can extract from Tane's methodology apply in other domains too:
A documented plan: A structured plan document acts as 'shared mutable stateshared mutable state' between human and AI. It is more systematic and easier to review than chat.
Stage-by-stage approval: No stage moves on to the next without the human's explicit approval. This constrains the AI's autonomous judgment and keeps humans in control.
Preserving context: Even in long sessions, the plan document preserves the essential context, so there's no need to dig back through the conversation history.
In the future, this pattern of collaboration will become the standard. The more powerful AI grows, the more the human role will shift toward the strategic and the judgmental. What matters is seeing AI not as a mere tool but as a collaborative partner with a clear division of roles. Tane's methodology lays out a blueprint for this future of collaboration.
