Boris Tane, head of the Cloudflare Workers Observability team, shared his experience using Claude Code.

His entire workflow follows a strict rule: before reviewing and approving a written plan, Claude is not allowed to modify a single line of code.
Most people use it by: providing a prompt, reviewing the output, fixing errors, and trying again.
This cycle works for simple tasks, but for complex features, it can take three times longer to clean up the mess.
The real problem isn't that Claude makes syntax errors; it's that before you confirm the assumptions, it has already spent fifteen minutes heading in the wrong direction.
His approach is: first, have Claude deeply read the codebase and write inline comments—not explanations in chat, but directly marking the wrong parts with "Incorrect, change to this."
This cycle is repeated one to six times until he's satisfied, then he says, "Implement it all."
The article continues: